Andrew Ng has been signaling this for a year. Amol Avasare recently doubled down on Lenny’s Podcast. The consensus? AI automates the low-leverage parts of the Product Management job, it amplifies the need for the higher order, more human parts of the job.

I spent the last few weeks in deep dives with leaders at Google, Meta, and leading Financial Services and Healthcare. These are the people building the agentic age. Their takeaway is that the “hustle” of managing feature complexity is becoming a commodity. What remains is the only thing that pays: Judgment.

The 2026 Reality: Rebuilding “Good PM / Bad PM”

AI is a force multiplier for Product Management judgment, not a crutch for answers.

Swipe Through: The AI PM Playbook

Resources

The history of personal productivity for people seeking high leverage out of their day is a cycle of failed attempts to automate human action without regard for taste. In today’s world, most of the Product leaders that I talk to are talking about creating a Personal OS. A force multiplier that helps them maximize their impact while minimizing the headaches of modern life.

Back in the early 2000s, as people started to get their first, second and third email addresses from AOL, CompuServe, their school, or work, their email contacts were spread out across a variety of clients and services. Plaxo, a service by Sean Parker, achieved one of the first true viral loops by promising to solve the fragmentation of personal contact management.

The early dream of automated contact management integrated directly into the workflow. Source: Scott Hanselman

They were mostly successful in aggregating an initial personal contacts database across silos of school, work, and some personal email services. The talk around the valley among leading technologists shifted to creating a Personal Relationship Manager (PRM) on top of this new wealth of data.

Ostensibly, the goal was to create leverage and social capital by automating the work of remembering birthdays and tracking news. While the idea was viral, the value was hollow. Automated messages felt scripted because they prioritized the completion of the task over the authenticity of the connection. The era of Plaxo and the PRM taught us that blind aggregation and automation without authenticity and taste turns connection into spam.

The Context Fragmentation Tax

The SoLoMo (Social Local Mobile) era, a term coined by John Doerr, put the world in our pockets. The mobile revolution connected everyone to each other while simultaneously establishing the technical infrastructure for daily digital existence. However, it perpetuated a new problem where data became trapped in fragmented stores like iMessage, Slack, and Email. We solved the problem of connecting people and devices but we created the problem of disconnected data trapped within individual apps.

Today, in the absence of a unified context layer, high-leverage leaders who want to maximize the impact of AI face five structural hurdles:

1. The Manual Routing Tax: Users act as information janitors, manually copying data between silos to maintain a coherent view of reality. This is a functional friction that wastes high-leverage time.

2. The Privacy Trade-Off: Users are forced to choose between utility and privacy. You can use integrated cloud ecosystems that offer magical convenience but require total data exposure, or you can use encrypted silos that are secure but “dumb” because they can’t talk to each other. There is currently no middle ground that offers both intelligence and sovereignty.

3. Context Amnesia: Siloed data has no memory. Every time you switch apps, your digital brain resets to zero. You have to re-explain your context to every new tool because the system lacks a persistent memory layer.

4. Incentive Misalignment: Current apps are optimized for engagement (time on site), not utility (time saved). It is not strategically rational for most apps to freely share their context with an aggregator, as their business model depends on keeping you inside their walled garden.

5. The Legacy API Bottleneck: Industry attempts to solve this via APIs or protocols (like MCP) are slow to take hold. Real-world tasks—like pulling lab results from a patient portal—often rely on legacy systems that will never offer clean APIs for a user to access programmatically.

For the power user looking to create leverage, operating within these constraints is a legacy state that talent should outgrow. This demand has given rise to the concept of a Personal OS—a unified intelligence layer that sits above your apps to extract signal from the noise.

ClawdBot Moltbot represents a Fundamental Architectural Shift

The emergence of Agents like ClawdBot Moltbot that create unified context represents the first architectural shift designed to solve this context routing issue. The ClawdBot Moltbot project went viral in technical circles by promising to bridge these gaps. Developed by Peter Steinberger, this open-source agent allows all of a user’s context—work, personal, and team—to be looked at holistically. (Note: During the drafting of this post, the project was renamed from ClawdBot to Moltbot).

Early adopters identify the transition from a 'personal assistant' to a proactive context-persistent tool. Source: @danpeguine

Moltbot accesses “raw metal” to bridge silos, unbound by any single vendor or API. It can run locally or on a private VPS, giving the user control over where their data lives.

Two interesting notes from this are that people are running ClawdBot Moltbot on Mac minis rather than a $5 VPS. My thesis is that this represents a deep-rooted user desire for privacy. While the Overton window has shifted a little bit in terms of social privacy and what people are willing to share on the web, there is an inherent desire to keep some information private.

Second interesting note, while this maybe an artifact of the early adopter community, is the user’s desire for Ownership over the digital means of production. This ownership is a strategic imperative. If you don’t own the data graph, you cannot compound its value. It’s perhaps the realization of the strategic value of switching cost created by time. This is the Lindy Effect applied to personal context. As the agent ages, it becomes more valuable. Once an agent knows a deal history and decision logic spanning a decade, it becomes an irreplaceable asset. That the early adopters are choosing to put this on a local machine that’s completely under their control signals a shift toward sovereignty.

For the Personal OS to become a reality, the industry must solve the critical problems identified by Moltbot’s early adoption

While the desire for a unified context is clear, the path forward is unpaved. The success of Moltbot proves that users are willing to tolerate friction to gain agency, but for this to scale beyond the technical elite, we must resolve four fundamental issues.

1. Table Stakes: Friction, Sovereignty, and Security The first set of issues are foundational table stakes. First, the setup barrier must be eliminated; a normal user will never configure a local server or manage API keys. Second, we cannot trade the “many small jails” of app silos for the “one big jail” of a vendor-locked Personal Context Cloud. Users need data sovereignty—the ability to own, export, and control their context graph. Finally, the “Open Shell” model of current local agents is a security non-starter. The industry must standardize a secure perimeter that authorizes an agent’s intent without giving it a blank check to every file on the system.

2. The Real Challenge: Expansion over Mimicry Once the foundation is secure, the true strategic challenge is style. Current AI models regress to the mean, producing generic prose that sounds like everyone else. A Personal OS that simply mimics your past behavior traps you in your own history.

To act as a true force multiplier, and not just go the way of Plaxo and the PRM, the agent must enable Expansion. It should not just replicate how you have written; it should help you express how you intend to write. I once lost a rental deposit because I didn’t fully know my rights. An expansion-focused agent would have flagged that knowledge gap instantly, using my context to arm me with the specific legal leverage I lacked. The goal is not to outsource and automate your voice away. Your personal OS should amplify your full potential and help you be all that you can be, rather than just a faster version of who you were yesterday.

3. The Vanity Trap If we fail to solve for expansion, we fall into the trap of Vanity Productivity. This is the risk of mistaking the feeling of generated copious output and iterations with the agent for achieving the reality of a strategic outcome. The real risk is becoming irrelevant because your competitor’s agent is negotiating better deals than you are. Without a focus on leveraging human potential, the Personal OS becomes an engine for generating noise, allowing users to feel productive while achieving nothing. The metric of success must shift from “actions taken” to “leverage gained.”

The Strategic Payoff

The objective is to make the user more present, allowing them to fully express themselves in ways never before possible rather than simply becoming a just more productive cog in a wheel. ClawdBot Moltbot isn’t just an app; it is the first credible blueprint for this shift from manual routing to personal sovereignty.

And once we solve this for ourselves, the next frontier is what happens when our agents start talking to each other. How are you architecting your own digital factory? Are you building an irreplaceable moat, or just adding to the noise?

Join the discussion on LinkedIn.

Last month, I outlined the four pillars of the current longevity stack. While that framework is useful for categorization, the real value lies in how these components integrate to solve specific clinical bottlenecks.

Today, I want to look at a specific case study: cardiovascular health. The field is moving from a model of simple lipid management toward “debugging” the underlying biology and architecting the biomimetic repair infrastructure required for true longevity.

The starting point is a major review published in Nature Reviews Cardiology (2026), which identifies a critical challenge: “Residual Cardiovascular Risk.”

The paper highlights a critical nuance. While aggressive lipid management has been a triumph of modern medicine, it has reached a point of diminishing returns. Even with optimal care, patients face a persistent risk driven by other factors, specifically, inflammatory pathways.

This clinical finding connects directly to Pillar 1 (Foundational Science). If inflammation is the signal identifying this residual risk, the next logical step is to understand the biological mechanism generating it. One area where that foundational science is moving forward at a steady pace is the study of cellular senescence.

What is Senescence?

In the context of biological aging, cellular senescence is a state where damaged cells stop dividing but refuse to die. Instead of being cleared by the immune system, they linger in the tissue as “zombie” nodes.

Diagram illustrating the transition from healthy cells to senescent ‘zombie’ cells and their impact on surrounding tissue. Source: National Institute on Aging (NIH)

While this mechanism is a critical tumor-suppressor in early life, its accumulation in aged tissue becomes a primary driver of decay. This is a classic example of antagonistic pleiotropy (Campisi, 2005), the principle where a mechanism that is beneficial in early life becomes detrimental as the organism ages.

This “zombie cell” accumulation is a systemic challenge, impacting tissues from the lungs to the kidneys. While I am focusing on the cardiovascular implications today, the broader impact of senescence across the four pillars is something I will explore in future posts.

Within Pillar 1, researchers are moving past identification toward active intervention. There is steady progress in the ability to detect and address these drivers of decline. This transition from decoding biology to engineering repair defines the next era of longevity.

Managing Residual Risk

The practice of medicine has seen progress in managing the most visible drivers of cardiovascular decay. Over the last few decades, Major Adverse Cardiovascular Events (MACE) have come down, driven largely by the widespread adoption of statins and, more recently, modern PCSK9 inhibitors.

A recent meta-analysis of eleven clinical trials involving over 135,000 patients led by Dr. Alberto Cordero (Cordero et al., 2023) established that intensive lipid-lowering therapies reduce the risk of MACE by an average of 15% and cardiovascular mortality by 6%. Building on this, clinical trials such as FOURIER (Sabatine et al., 2017) and ODYSSEY Outcomes (Schwartz et al., 2018) have shown that modern PCSK9 inhibitors can reduce LDL cholesterol by 50-70% and further lower the risk of MACE by an additional 15-20% in high-risk populations.

The blue zones illustrate the era of rapid lipid innovation and the resulting decline in mortality. The graph shows the performance plateau and stalling progress described in Nature Reviews Cardiology. Source: National Center for Health Statistics (CDC)

Researchers suggest the answer to this stagnation lies in “residual inflammatory risk.” As established in The Lancet (2023), inflammation, specifically as measured by high-sensitivity C-reactive protein (hsCRP), can be a stronger predictor of future cardiovascular events and all-cause mortality than LDL cholesterol levels for patients already receiving statin therapy.

This suggests that the next frontier is not about managing lipids, but addressing the aging of the vascular tissue itself.

The SASP: A Biological Broadcast Storm

A major driver behind this elevated hsCRP signal is the Senescence-Associated Secretory Phenotype (SASP).

To understand the SASP, it is helpful to use a universal engineering analogy: a broadcast storm. Imagine a malfunctioning Ethernet node or a “stuck mic” on a radio frequency. These senescent cells don’t simply stop functioning. They begin broadcasting a constant stream of pro-inflammatory signals, including cytokines like IL-6, which triggers the liver to produce CRP, chemokines, and matrix metalloproteinases (MMPs). This noise eventually drowns out healthy cellular intent, polluting the environment and degrading the structural integrity of the heart and arteries. This is a core driver of myocardial fibrosis and arterial stiffening.

Reconstructing the Biological Infrastructure

Clearing the broadcast storm of the SASP via senolytics is a necessary first step, but it is not enough for longevity improvements. Removing the malfunctioning nodes stops the active signal pollution, but it does not automatically resolve the accumulated structural debt, which is the stiff, non-conductive scar tissue left behind by years of chronic inflammation.

Moving from cleanup to active repair requires a solution for the lack of organized cellular housing. Professor Onnik Agbulut and his team at Sorbonne University’s Institute of Biology Paris-Seine are addressing this by engineering bio-inspired biomaterials. These synthetic structures, including 3D scaffolds and electrospun nanofibers, function as a surrogate for the extracellular matrix. By providing this environment (Kitsara, Agbulut, et al., 2023), the Sorbonne team is creating the structural template required for new, healthy cells, such as human induced pluripotent stem cell-derived cardiomyocytes, to colonize and restore functional capacity to the heart.

The Roadmap Forward

The focus in cardiovascular health is shifting from simple mitigation to a roadmap of active biological repair. From a technologist’s perspective, the critical path to making these therapies a reality is the implementation reality of clinical delivery. The challenge is moving from successful in-vitro scaffolding to safe, scalable human applications.

The roadmap for these advances looks like this:

1. Quieting the Noise: Validating targeted senolytic interventions that can resolve the inflammatory burden of the SASP without systemic toxicity.
2. Repair Infrastructure Deployment: Engineering the delivery mechanisms for biomimetic scaffolds to ensure they can host and mature functional cardiomyocytes within a living heart.

By addressing the residual inflammatory risk and structural debt at their source, the next generation of therapeutics will move beyond patching a legacy architecture. This transition from “managing decline” to “architecting repair” is the shift required to move the needle on human longevity. The field is learning to reinforce biological integrity at the cellular level.

References

• Campisi, J. (2005). Senescent cells, tumor suppression, and organismal aging: good citizens, bad neighbors. Cell, 120(4), 513-522.
• Cordero, A., et al. (2023). The efficacy of intensive lipid-lowering therapies on the reduction of LDLc and of major cardiovascular events. Journal of Clinical Lipidology, 17(4), Supplement, S1-S2.
• Sabatine, M. S., et al. (2017). Evolocumab and Clinical Outcomes in Patients with Cardiovascular Disease. New England Journal of Medicine, 376(18), 1713-1722.
• Schwartz, G. G., et al. (2018). Alirocumab and Cardiovascular Outcomes after Acute Coronary Syndrome. New England Journal of Medicine, 379(22), 2097-2107.
• Ridker, P.M., et al. (2023). Inflammation and cholesterol as predictors of cardiovascular events among patients receiving statin therapy: a collaborative analysis of three randomised trials. The Lancet.
• Ottaviani, A., et al. (2024). Mechanisms of Telomere-Driven Senescence and Immune Evasion. IRCAN Research Perspectives.
• Kitsara, M., Kontziampasis, D., Agbulut, O., & Chen, Y. (2023). Surfaces for hearts: Establishing the optimum plasma surface engineering methodology on polystyrene for cardiac cell engineering. Applied Surface Science, 620, 156822.
• Batoumeni, V., Agbulut, O., et al. (2025). Integrated phenotypic and transcriptomic characterization of desmin-related cardiomyopathy in hiPSC-derived cardiomyocytes. European Journal of Cell Biology.

Last month, I outlined the four pillars driving the longevity revolution. I expected them to mature over the coming year. I didn’t expect to see three of them validated in the first three weeks of January.

Recap: The Four Pillars

To ground this month’s updates, let’s look at the innovation stack I outlined in December:

1. Decoding Cellular Aging: Targeting the fundamental biology of decline (See my technical deep dive on advances in Pillar 1 foundational biology here).
2. Democratizing Access: Moving clinical-grade diagnostics into the consumer’s hands.
3. Personalization at Scale: Using AI to orchestrate unique, data-driven health plans.
4. Real-Time Adaptation: Moving from static plans to living, breathing feedback loops.

In just the last few weeks, we’ve seen significant moves that validate Pillar 3 (Agentic AI) and Pillar 4 (Real-Time Adaptation), along with a regulatory tailwind for Pillar 2 (Democratizing Access).

1. The Fight for the Health Interface (Pillar 3)

We are seeing a push to own the consumer health interface. This isn’t speculative; it’s a response to a shift in human behavior. According to an OpenAI report from January 6, 40 million people a day are already using AI to get answers about their symptoms and healthcare coverage.

Data from a Knit survey commissioned by OpenAI showing how US adults are using AI for healthcare. Source: OpenAI

With OpenAI’s ChatGPT Health, Anthropic’s clinical integrations, Amazon One Medical’s Health AI, and Google’s Personal Health Agent, the industry is solving the “last mile” problem of health data. These platforms are bridging the gap between “siloed data” and “actionable insights” by aggregating records from lab results (HealthEx/Function) to wearable data (Apple Health/Google Fitbit) and pharmacy records.

This is the democratization of data in action. For years, your Electronic Health Record (EHR) was something you visited, not something you utilized. Now, that data flows directly into the models you use daily, capable of explaining test results in plain language and managing medication renewals in real-time. All these players are adopting a privacy-first approach, stating they will not train models on this private health data. This isn’t benevolence; it’s a market requirement. Without this guarantee, the “consumer health” play is dead on arrival.

While they are competing for the consumer front-end, Anthropic is also tackling the “unsexy” backend. Their expansion into healthcare operations, which includes automating prior authorizations, claims processing, and clinical trial protocols, is the infrastructure layer that makes longevity scalable. By reducing the friction that currently strangles clinical innovation, we accelerate the entire pipeline. Faster trials mean faster therapeutics.

The Insight: The combination of consumer data democratization and backend operational efficiency means the industry is building the “structural rails” for a system that can handle personalized care at scale.

2. From Measurement to Prediction (Pillar 4)

I’ve written before about Making Wearables Useful, focusing on how the initial novelty of seeing “latent data” (like your step count or glucose level) eventually wears off. We are evolving past that phase.

Abbott’s new “Libre Assist” feature is a clear example. Using generative AI, it analyzes your meal photo to predict the spike before it happens, moving beyond simple after-the-fact reporting.

The Abbott Libre Assist interface showing a "Major impact" prediction for a chicken dinner, along with real-time adaptation tips. Source: Abbott

The Insight: This changes the user psychology. Instead of feeling guilty about a spike that already happened, you get a moment of agency. I might still choose to eat that celebratory donut, but knowing the predicted impact, I can plan for a 15-minute walk shortly after to blunt the curve. That shift, moving from retroactive guilt to proactive planning, is the essence of Real-Time Adaptation. It shrinks the feedback loop from two hours (the post-meal spike) to zero minutes (the pre-meal choice).

3. The Regulatory Floodgates Open (Pillar 2)

The biggest bottleneck for health innovation hasn’t been technology, it’s been regulatory ambiguity. That’s why the FDA’s recent announcement (Jan 2026) is so significant.

FDA Commissioner Dr. Marty Makary announced that the FDA will not regulate general wellness products, including devices that measure biomarkers like blood pressure or glucose if they are used for nutritional or lifestyle choices rather than disease management.

The Insight: This “opens the gates” for AI models to predict biomarkers in healthy populations using novel sensors. We can now deploy technologies like PPG to predict blood pressure (without the cuff) or non-invasive glucose monitoring for wellness without the regulatory burden of a Class II medical device.

This creates a lane for the “Quantified Self” to mature into “Quantified Health.” This shifts the risk and the power to the consumer. It allows us to build a “pre-clinical” layer of health data without being strangled by red tape intended for critical care.

The Bottom Line

Structural rails are being laid. The industry now provides personalization (OpenAI/Anthropic/Amazon/Google), prediction (Abbott), and access (FDA). The tools to engineer a personalized healthspan are no longer science fiction; they are here, and they are scaling.

My career has been a hunt for force multipliers. I fell in love with engineering because software is the ultimate lever to scale human impact. I moved into leadership to scale that further, orchestrating teams to solve problems no individual could tackle alone.

Update (January 2026): I recently published a deep dive into how these pillars are being validated by the latest moves from Abbott, Anthropic, and the FDA. You can read the January 2026 Longevity Update here.

I’ve long been a proponent of the Quantified Self movement and the ability to understand what’s happening in our bodies to maximize performance. For years, however, we were largely measuring the periphery of the problem, focusing on performance and tracking metrics such as steps, sleep scores, HRV, or resting heart rate, without a clear eye toward holistic system-level optimization.

We have hit an inflection point where we are starting to understand the core of the system: resilience. When we examine quantified self data through this lens, we quickly arrive at the interconnected ideas of aging, longevity, and healthspan.

For too long, we accepted that performance, both individual markers and systemic output, inevitably declines with age. That as your biological age increases your resilience decreases. Now, new views are emerging that Aging should be a looked at a cluster of symptoms, if not a diseases. There is historic precedence here. We used to define Cancer as a single disease but no oncologist considers all the various cancers to be one disease or attempts to treat them all the same way.

For too long, we accepted that performance, both individual markers and systemic output, inevitably decline with age. We assumed that as biological age increases, resilience decreases. Now, new perspectives are emerging: that what we call aging should be viewed as a cluster of symptoms, if not multiple diseases. There is historical precedence here; we once defined cancer as a single disease and potentailly with single origin, but today no oncologist considers all of cancer’s various forms to be one disease or attempts to treat them all the same way.

We’re at an intersection where we will start move simply tracking the decline to actively engineering our capacity for sustained, high performance based on rigorous, scientific pursuit of a future where healthspan is engineered, not just inherited.

Below are the seismic shifts underway that make me excited about the longevity industry and some of the critical challenges we must solve to make this available to all people.

Four Key Trends in Longevity

1. Foundational Science is Decoding Cellular Aging

For decades, aging was treated as a black box, an inevitable decline managed only by treating symptoms as they appeared. That era is ending. We are now decoding the core Hallmarks of Aging, moving from guesswork to targeting specific biological mechanisms. Recent breakthroughs have illuminated drivers like cellular senescence (the accumulation of ‘zombie cells’), mitochondrial dysfunction, and epigenetic alterations. For instance, researchers have pinpointed specific mechanisms such as the role of high-molecular-mass hyaluronic acid (Tian et al., Nature 2013) and cGAS-mediated DNA repair (Chen et al., Science 2024) in the exceptional longevity of naked mole-rats. These discoveries are shifting the paradigm from managing decline to engineering resilience at the cellular level.

2. Clinical-Grade Diagnostics are Becoming Widely Accessible

Vital health signals were once locked in clinical labs or trapped in device silos. For years, data like HRV (Heart Rate Variability) from devices like the Apple Watch remained inaccessible and unactionable for the user, while critical clinical markers such as ApoB or hsCRP required a specialist’s prescription. Now, however, those walls are coming down. Consumer-facing panels from companies like Function Health provide direct access to clinical markers, and data from wearables and CGMs (from Abbott and Dexcom) is becoming increasingly open and interoperable. This shift allows individuals to see the same high-fidelity picture of their health once reserved for a research setting.

3. Cost-Effective Personalization at Scale

True personalization has always been an economic problem. Creating a cohesive health plan that integrates nutrition, exercise, and medical history required a “high-touch” team of expensive experts, including a doctor, nutritionist, and trainer, to manually coordinate and synthesize data. This model was an unscalable luxury. Agentic AI is dismantling this barrier. As outlined in research on Google’s Personal Health Agent, LLMs can now reason across diverse data streams, from sleep patterns to medical records, to generate personalized insights. The agent effectively acts as an affordable, scalable “expert team,” democratizing the level of coordinated care that used to cost thousands of dollars a month.

4. Real-Time Adaptation and Continuous Care

The traditional “personalized health plan” suffered from a fatal flaw: it was static. Whether delivered as a PDF or a consultation, the advice was fixed and immediately outdated by the realities of daily life, such as stress, travel, or injury. We are moving to a model of continuous care where the plan is alive. The new era of AI assistants provides dynamic adaptation in real-time. An agent can detect a poor night’s sleep and automatically lower your workout intensity, or modify your nutrition plan based on a glucose spike from lunch. This creates an unbroken feedback loop that adapts to your biology in the moment, transforming health from a periodic check-in to a continuous, proactive activity that helps you meet the moment.

The Towering Challenges Ahead

This progress is real, but the hurdles are immense. Simply having the technology is not enough.

1. The Mindset Challenge: Moving from Luxury to Necessity

The biggest barrier isn’t scientific; it’s cultural. For most people, proactive health is still seen as a “vitamin” or a luxury, not a critical “painkiller.” We need to shift the Overton Window, just as we did for heart disease and strokes. We came to understand that these weren’t just random acts of fate but diseases that could be prevented. We must frame chronic disease not as a certainty of aging, but as a issue that we can address and something that is to be priortized at individual level if not a the population health level.

2. The Evidence Challenge: From Anecdote to Science

The wellness space is a noisy landscape, and it can be difficult to separate credible science from marketing hype. People are already exploring holistic medicine, contrast therapy, and peptides, often relying on anecdotal evidence and questionable influencers with broad reach. Rather than dismiss this curiosity, we should empower it with better tools. Better tools to evaluate the claims and make it accessible to consumers to understand of the levels of scientific proof that has been estabilished.

Another aspect to consider is a personalized experimental platform to help individuals correlate their own data with the actions they take. Want to figure out why you have headaches on Wednesday - knowthing that you run and then do hot yoga and only drink 8oz of water on Tuesday might be helpful. While this is not a replacement for rigorous science and cannot prove causality, it provides a crucial feedback loop. It gives people the tools to move from anecdote to personal insight, helping them understand what’s actually working for their unique biology. With some controls, these personalized platforms could lead to population level insights like the recent article from Oura on the impact of Alcohol and Sleep

3. The Monetization Challenge: Who Pays for Not Getting Sick?

This is the most critical systems-level problem. Our economy has robust financial rails for ‘sick care,’ but almost no infrastructure to support preventative therapies at scale. We’ve managed to fund effective interventions like smoking cessation and diabetes prevention programs, yet we lack a population-level funding mechanism for the next generation of proactive health. The emerging concierge models are promising but exclude the vast majority. The crucial, unanswered question is: How do we build the business models that make preventative health accessible to all, not just the wealthy?

I remain incredibly excited about the opportunity here. The challenges are significant, but the potential to fundamentally change population health is even greater. I’m looking to connect with other leaders, builders, and thinkers who are working to solve these problems. If you are building something meaningful in this space, let’s talk.

Part 3 of the AI Exploration Series. Missed the earlier posts? Start with The Spark of an Idea or Giving the Ghost a Machine.

What started as a fun AI experiment quickly became a lesson in responsibility. The “Uncleji” project, a generative AI persona designed to mimic the quirks and wisdom of a beloved uncle, showed us that moving from sandbox to production with non-deterministic AI isn’t just about building cool tech; it’s about safeguarding trust, user safety, and controlling unseen costs. The “move fast and break things” mantra of Web 2.0 is lethal in this new era. Here are three lessons we learned about building a system that deserves user trust.

1. The Financial Firewall

Before a single public user interacted with our agent, we realized that a public-facing LLM is a target. The risk isn’t just a service outage; it is compute hijacking. If an attacker bypasses defenses, they can exploit the API to run their own workloads, potentially leaving us with a catastrophic bill.

We had to design a defense that assumed attacks would happen.

• Cost Control & Safeguards: We configured Google Cloud Billing Budgets and API quotas immediately. These act as an automatic circuit breaker, alerting us or pausing services if usage spikes unexpectedly.
• Edge Defense (Cloud Armor): Leveraging tools like Google Cloud Armor helps block malicious traffic and botnets before they ever reach the application.
• Application Defense (Redis & Signatures): We applied standard web security practices that are often overlooked in AI demos. Verifying HMAC signatures prevents tampering and replay attacks. Furthermore, enforcing distributed rate limits via Redis ensures that even as our server instances scale up to handle traffic, no single user can bypass quotas by hopping between containers.

This multi-layered approach ensures the financial and operational integrity of the system remains secure.

2. The Bartender Paradox

To make an agent feel “real,” it often needs short-term memory. An AI that forgets you every two seconds is useless. But memory isn’t one-size-fits-all. A medical AI needs a clinical history; a travel agent AI needs your dates; a casual chat bot just needs the “vibe.”

The challenge is balancing this utility with privacy. What are we remembering, and about whom?

We adopted a framework I call the “Bartender Principle.” A great bartender knows your usual drink and remembers your last conversation (“How did that meeting go?”), but they don’t need your social security number or your home address to do it. They offer recognition without intrusion.

For “Uncleji,” this meant implementing pseudonymized user IDs (hashing user identifiers). This enables session memory—so the uncle remembers you asked about a recipe five minutes ago—without storing Personally Identifiable Information (PII) like your actual phone number. It balances the human need for connection with the ethical need for anonymity.

3. Semantic Observability

In traditional software, we monitor for latency and uptime. In AI, we realized we must monitor for intent.

Even if “happy path” testing was green—the persona was funny, retrieval was accurate—the gap often lay in operationalizing abuse-specific observability.

• Traditional Observability: Is the server up?
• AI Observability: Is the AI being manipulated into hate speech? Is it hallucinating dangerous advice?

True production readiness required leveraging tools (like Langfuse) not just for debugging code, but for analyzing semantic traces at scale. We needed eyes on the conversation, not just the console.

Earning User Trust

The Uncleji project proved that innovation without guardrails is just liability.

As builders, we must ask: Are we building cool tech, or are we building trusted systems? The most essential user needs haven’t changed since the dawn of commerce: safety, privacy, and reliability. These are the foundations of the contract we make with our users.

In the age of AI, our code determines what the product can do. Our guardrails determine what it should do.

Giving the Ghost a Machine: The Art of Engineering for Creative AI Services

Part 2 of the AI Exploration Series. Read Part 1: The Spark of an Idea.

In our last post, we explored the spark of an idea: harnessing AI’s “hallucinations” as a feature. But a creative concept—a ghost—is ephemeral just a demo. To bring it to life and share with real users to let it interact with the world, it needs a body. It needs a machine.

This is the story of building that machine for Uncleji. It’s not about enterprise-grade software development and corporate compliance in the traditional sense but we need to bring that mind set to protect both AI infrastructure and the Users while ensuring the work is resilient and accessible.

The architecture bridges robust cloud infrastructure with modern AI tooling:

Vibe Coding ≠ Skipping Best Practices

We have over 10 years of tried and trued best practices for deploying web services using cloud infrastructure. We can and should re-use these in deploying AI services like UncleJi.

• Containerization for Consistency (Podman & Google Cloud Run): The first step was to ensure Uncleji behaved predictably, whether on my machine or on a server for thousands to interact with. Using containers is like creating a perfect, climate-controlled gallery case for the art piece. It ensures the experience—the art—is identical for every viewer, preventing the sculpture from crumbling the moment it leaves the studio.

• Security Best Practices: The speed of “vibe coding” often invites sloppy security, like hardcoding API keys. But moving to production demands enterprise standards. We utilized Google Secret Manager to ensure credentials remained secure, protecting the “magic” from misuse without slowing down development.

• Traditional Observability for Foundational Health: Before delving into LLM-specific tracing, the fundamental health of Uncleji’s “machine” relies on robust cloud observability. Built-in Google Cloud tools like Cloud Monitoring and Cloud Logging were crucial for detecting abuse patterns, system outages, and performance bottlenecks—just as we would for any critical cloud-based deployment.

Infrastructure as a Conversation: The best part is that you don’t need to be a DevOps wizard to build this foundation. I used the Gemini CLI to spin up this infrastructure. I effectively treated the infrastructure setup as just another conversation. By leveraging the Gemini CLI, the gcloud CLI, or specifically the Google Cloud Run MCP, you can simply ask for what the best practices and pitfalls are in deploying your service, and work with the LLM to generate a plan for a production-ready environment and let it handle the plumbing.

This foundation ensures the art can be shared without being compromised. It creates a reliable stage where the performance can happen.

Tuning the Ghost’s Voice: The Artist’s Control Panel

With a stable vessel built, the next challenge was LLM behavioral observability. Standard cloud logs are essential for uptime, but they are blind to the nuances of AI persona and response quality. You can’t just grep for “personality drift.”

To navigate this new frontier, we must embrace specialized tooling designed for the probabilistic nature of LLMs. I integrated Langfuse not just for tracing, but as a dynamic control plane.

Moving the system prompts out of the codebase and into Langfuse was a game-changer. It allowed me to debug and iterate rapidly on Uncleji’s persona—understanding where the model was deviating and tuning his “creative” responses in real-time without the friction of a full commit-build-deploy cycle.

Anecdote: Taming Uncleji’s Ramblings

The data was clear: users disengaged during long monologues. But instead of a week-long engineering sprint to rewrite code, I simply adjusted the “knobs” in Langfuse.

Instead of filing a ticket to engineering (also me) to tweak the prompt in code, I went directly into Langfuse. It was like being a musician at a mixing board. I could adjust the “knobs” on his personality, tuning the prompt to enforce shorter, text-message-sized chunks and more frequent conversational turns.

The result was immediate. User engagement improved, and the feedback shifted from “he talks too much” to “I love his stories!” This change required zero engineering effort for the adjustment itself. It was a moment of pure, creative flow, where the tool got out of the way and allowed the artist to directly shape the creation.

The Machine in Service of the Ghost

It’s easy to get captivated by the magic of LLMs or the new capabilities of tools like Langfuse. But innovation cannot be an excuse to abandon first principles. A creative AI that is unstable, insecure, or unobservable is ultimately a liability, not an asset. Robust, boring engineering is the most powerful enabler of creative AI.

By building a stable “vessel” with proven infrastructure and pairing it with a dynamic “control panel” for the persona, we give the “ghost” the freedom to be truly expressive. The disciplined craft of engineering provides the stable canvas on which the art can finally come alive.

In our final post, we’ll explore what happens when you release this creation into the world—and the surprising lessons it teaches us about trust, safety, and the human-AI connection.

Next Up: How do you protect your creation and your users once it’s live? Read Part 3: Responsible AI by Design: Leadership Beyond the Code.

The Spark of an Idea: Turning AI Hallucinations into a Product Feature

The AI conversation has been dominated by a single obsession: eliminating ‘hallucinations.’ We’re told they are a bug to squash, a flaw to engineer out. But this always felt like a failure of imagination.

So, I started a deliberate experiment with a different question at its core: What happens if, instead of fighting AI’s unpredictability, we harness it as a feature? This is the story of that exploration, starting with an AI named Uncleji, which ran as a live production experiment over the summer.

From Vibe Coding to Generative Identity Design

I began my journey with Uncleji, an AI conversational agent embodying a charming, meandering Indian uncle. The initial work was a process of generating an iconic profile image to get the right visual feel.

But this experiment evolved into a foundational practice I call Generative Identity Design.

This is a shift from scripting a bot to co-authoring a persona. Instead of meticulously writing every line of dialogue, we engaged in a deliberate loop where the AI—guided by high-level constraints—helped discover its own backstory, limitations, and voice.

This allowed us to rapidly integrate add new context to the personality, iterate on conversational flow and incorporate user feedback into the core of the character rather than patching individual responses. When early testers felt Uncleji needed to feel more dynamic without losing his charm, we tuned the prompt to encourage:

• Concise Responses: Limiting his output to short, punchy phrases to align with modern chat etiquette versus the verbose LLM responses.

• Contextual Slang Integration: Incorporating relevant Gen Z slang, but with an authentic, non-forced feel, ensuring he didn’t sound like a “typical chatbot.”

• Proactive Conversation Flow: Designing the prompt to encourage him to ask questions and “keep moving the conversation ahead,” fostering engaging dialogue unlike a dry Q&A or a rote lecture.

This process turned prompt engineering into character direction, bringing Uncleji’s persona to life.

A new rapid prototyping skill

Uncleji was the test case, but the implication is broader. This method offers a new form of rapid prototyping.

The paradigm for building AI products is fundamentally different. Traditionally, we build software by meticulously outlining previously know requirements and coding its behavior. With AI, we are increasingly designing the system that generates the product’s behavior and constantly tuning it to meet the needs.

This creative process, demonstrated here with Uncleji, is a form of rapid prototyping applicable to any AI product. Before building complex infrastructure, your teams can use high-level prompts to define and test an AI’s core logic—be it a conversational style, a coding methodology, or a data analysis framework.

Beyond the Prototype: A Playbook for Unconventional AI

This initial exploration into the spark of an idea—using AI hallucinations as a creative tool—revealed that moving from a “vibe code” to a production-ready system requires a deliberate, structured approach.

However, taking Uncleji from a local experiment to a public art piece introduced fascinating new creative constraints beyond just the “vibe”:

• Safety: How do you let an AI be wildly creative while ensuring it is never harmful?
• Cohesion: When “accuracy” is irrelevant, how do you measure if the AI is staying true to its character?
• Trust: How do you build rapport with an audience when your narrator is intentionally unreliable?

We had the spark. We had the persona. But a creative concept running on a local machine is just a ghost. To bring it to life—scalably, safely, securely, and reliably—it needed a body.

Read Part 2: Giving the Ghost a Machine.

Context Pollution - A Natural Outcome of Poor UX

I was vibing an hour deep into a session with one of today’s most advanced AI agents, trying to hammer out a new idea exploration. We were pulling in industry reports, summarizing articles, and iterating on multiple sections of the document at once. In that flow state, the AI felt like a true partner.

Then suddenly, things started getting wonky. I had to repeat prior instructions. The agent started reintroducing concepts we’ve dropped 30 mins ago. I scrolled up to see what’s happening.

It was a sprawling, chaotic battlefield of ideas. We had easily crushed over 100,000 tokens of back-and-forth. The brilliant final strategy was in there, but it was buried under mountains of discarded summaries, false starts, and abandoned tangents. And it hit me: the AI hadn’t been a partner in my creative process; it had been a court reporter, meticulously recording every word but understanding none of the intent.

This is a critical design failure I call Context Pollution. It’s what happens when an agent meticulously tracks every output, the literal transcript, while completely missing the outcome we are trying to achieve.

The High Cost of Context Pollution

Context Pollution is what happens when an agent’s history becomes so cluttered it degrades the quality of collaboration. The problem is concrete:

• Confusion & Drift: The agent loses the plot, referencing discarded ideas.
• Performance Degradation: It re-processes thousands of irrelevant tokens, slowing the creative process.
• Increased Cost: Every token costs money. Rereading a novel’s worth of brainstorming to add a sentence is fantastically inefficient.

Now you might be thinking, no problem, my model has a 1 Millon Token Context window that solves this. It doesn’t. A bigger window is just a bigger junk pile to search for the critical needle of right information. It simply postpones the inevitable signal loss.

AI Doesn’t Get vFinal_Final_revised_02.pdf

The root of this problem reveals a deeper truth: our current agents are literalists in a world of nuance. They don’t understand that the deep creative act of creation is a messy process. Their design misses a key point in how humans create: The process itself is iterative and deep BUT ephemeral. There are many intermediate outputs, but the outcome is what matters.

We’ve all seen the AI declare, “Here is the best and ultimate final version!” only for us to immediately reply, “That’s a good start, but change the tone.” This mirrors our own chaotic file habits. Our desktops are littered with Exec_Presentation_v2_final, Exec_Presentation_v3_final_new_final, and Exec_Presentation_v4_USE_THIS_ONE.

Source: X - @Studio_aaa

The difference is, we know which one is the ground truth. The AI does not.

The Emerging Art of Context Engineering

To combat the general issue of Context management, a new discipline is emerging: Context Engineering LangChain - Rise of Context Engineering. It is the art of skillfully structuring prompts and managing conversational history to guide an AI toward a desired outcome. A recent guide on Effective Context Engineering from Anthropic details the incredible effort required to do this well.

Source: Dex Horthy

This discipline includes several clever computational techniques:

• Automated Summarization: This technique, found in frameworks like LangChain, automatically condenses the conversation.
• Retrieval-Augmented Generation (RAG): This powerful tool connects an LLM to external knowledge, as explained in depth by Meta AI.
• Tree-of-Thought: This advanced prompting technique, detailed in the paper “Tree of Thoughts: Deliberate Problem Solving with Large Language Models,” encourages the LLM to internally brainstorm different reasoning paths.

While these techniques are powerful, they all frame the issue as a computational problem. They are trying to solve a human behavior problem with more sophisticated engineering, when a better understaing user needs might guide us to the simpler answer.

A New Model: The “Branch and Fold” Methodology

What if we solved this with design instead of just computation? I call this new design pattern the “Branch and Fold” Methodology.

The solution is to change the user interface from a linear log to a two-dimensional information plane. The main conversation flows vertically. But at any point, the user can create a horizontal Branch—a self-contained sandbox for exploration. As you can see in the prototype below, this isn’t just a separate chat; it’s an explicit ‘Brainstorming Active’ mode. This creates a self-contained sandbox where the exploration has its own dedicated session token count, completely isolated from the main conversatio

When you find the insight you need, you Fold the branch. The messy exploration collapses, committing only the distilled outcome back to the main conversation. The context pollution vanishes; the outcome remains.

The proof is in the pudding: the solution lies in a more thoughtful interface. To make this tangible, I started describing the concept to Gemini 3.0 in Google’s AI Studio, intending to build a simple UX mock-up. But something incredible happened: with the clarity of the user-centric concept I provided, the AI didn’t just create a mock-up; it built the basis of a full working prototype.

Here is The ‘Branch and Fold’ prototype in action. A sandboxed ‘Brainstorming Active’ session has its own self-contained token count, keeping the main conversation clean and focused with a collapse Artifact the synthesizes the information from the brainstorming.

This idea is built on core principles of good product management: matching technology to user intent, driving outcomes over outputs, and building trust through transparency. I am looking forward to refining and experimenting to quantify the impact.

The Tangible Benefits

This user-centric model delivers immediate benefits. It creates True Composability, allowing finalized branches to become reusable components. It Increases User Trust and Control, providing a “scratchpad” for free exploration. And it drives Efficiency Gains in quality, speed, and cost for any use case the model is tackling.

The Hard Questions & The Path Forward

The most pressing question is how a methodology designed for a visual UX translates to a non-visual interface like a CLI. This is a fascinating area because the CLI offers powerful workflow capabilities, like running agents in parallel, that consumer UIs can’t easily match.

In some sense, A developer’s workflow is already parallel; the opportunity is to make the AI a true parallel partner. Instead of today’s sandboxed manual context management tied to a specific branch, the system would dynamically manage the context as work is completed across the project, long before a final git merge. This moves beyond the current branch to a multi-branch awareness of the entire project—a profound step towards a true multi-agent development system. I am curious to see how this progresses.

Another big open question to contemplate is how this pattern could apply to agent-to-agent communication. How do Agents brainstorm amongst themselves and share interim outputs but also drive towards a shared outcome.

The Timeless Principle

Ultimately, the fundamentals of great product design haven’t been repealed by the AI revolution. It all comes back to a deep, obsessive understanding of how users think, work, and create. As Product executives our job is not to be mesmerized by the capabilities of this new technology, but to bend it to the needs of our users. The tools are new, but the mission is, and always has been, the same: start with the user.

“You need to be more strategic about how much you give.” Those words from my mentor hit hard as I moved from manager to leader. There I was, thinking my endless support and availability made me a great leader. Instead, I was sprinting toward burnout—and taking my team with me.

Here are three transformational books that redefined how I show up every day as a leader.

1. Give and Take (Adam Grant)

The book sets a framework for understanding the dynamics of “givers,” “takers,” and “matchers.” I was the quintessential “giver” Grant describes—always available, constantly helping, and slowly burning out. But here’s what changed everything: Learning that the most successful givers aren’t the ones who give the most—they’re the ones who give strategically. Another key breakthrough: You can’t pour from an empty cup. Setting boundaries doesn’t make you less generous—it makes your giving sustainable.

2. Crucial Conversations

“We need to talk” used to make my stomach drop. Now? It’s an opportunity for breakthrough. This book taught me that the most important conversations are about being building shared understanding and finding a path forward to discuss the C-P-R (Content/Pattern/Relationship) issue at hand. Real example: When a key team member seemed to be missing deadlines, being able to escalate to conversation about a repeated pattern helped us manage the tough conversation about workload and expectations. Result? They stepped up in ways I never expected, and I gained back 3 hours a week.

3. Dare to Lead (Brené Brown)

Brown’s research revealed that vulnerability is not a weakness as many have been taught to believe, but the foundation for authentic connection and trust. Embracing vulnerability and being ready to “rumble” as Brown calls it with thorny feelings and issues creates an opening for true partnership to emerge in any relationship.

The Real Breakthrough

These books go beyond leadership theory—they’re about human connection. They taught me that great leadership is about showing up authentically, setting clear boundaries, and creating space for others to do the same - not giving until you break.

Want your own breakthrough? Start with these books. But remember—the real transformation happens when you close the book and start having the conversations you’ve been avoiding. And that this is an ongoing process that you will get right and mess up on a regular basis. So be kind to yourself along the way too.

Share your leadership journey with me on LinkedIn. What conversation are you ready to have?

Picture this: It’s Monday morning, and your CEO wants to know when that game-changing feature will launch. Your engineering lead is talking about “technical unknowns.” Your designer needs “more discovery time.” And somewhere in your inbox, there’s a customer asking for the sixth time when they’ll see the update.

Sound familiar?

I’ve seen product launches implode because of vague timelines. One team was “targeting Q3” for a critical feature, which, when pressed, actually meant “maybe late September, if we’re lucky.” The result? Missed deadlines, frustrated customers, and a stressed-out team.

Want to know the difference between good product teams and great ones?

While everyone talks about accountability, great teams go beyond ownership and get precise. Knowing exactly how precise your timeline needs to be is the real secret weapon.

Product development and roadmaps often feel like navigating through fog. Outcomes are a blur, priorities shift, and getting predictable commitments feels like herding cats.

But what if I told you there’s a simple planning rule that cuts through this chaos?

Sandeep’s Rule for Planning: Precision is Power

The closer the deadline, the more precise your estimate must be.

It works like this, if your deadline is:

• 24+ months: Name the year
• 12-24 months: Specify the half-year
• 6-12 months: Pin down the quarter
• Within 6 months: Commit to the month
• Under 6 months: Lock in the date

This isn’t just another planning framework—it’s a mindset shift that transforms how teams approach timelines.

Why It Works

In my years of scaling products, I’ve watched teams hyperfocus on immediate deliverables while treating future milestones as abstract concepts.

Here’s the trap I see teams fall into: Labeling something as “next half” makes it feel safely distant. But if you’re starting that conversation in early May, reality hits different—you might have just four weeks to deliver, depending on how that timeline was understood. That psychological distance between “next half” and “four weeks” can be the difference between success and scrambling.

But here’s what’s fascinating: When we apply the Timeline Planning Rule to distant goals—even just specifying “H2 2026” instead of “future”—something clicks. Teams start thinking deeper. They spot dependencies. They uncover complexities. Most importantly, they begin taking actions today that pave the way for tomorrow’s success.

For example, when a team commits to a specific month, they’re forced to break down the project into smaller, more manageable tasks. This reveals dependencies and potential roadblocks early on. Instead of a vague “next half” project, they’re dealing with a concrete “September launch,” which triggers specific actions like “finalize design by July 15th” and “begin user testing by August 1st.”

Put It Into Practice

The Timeline Planning Rule does more than set dates—it builds a culture of clarity and accountability. It turns vague possibilities into concrete commitments. It transforms good product teams into great ones. It empowers teams to anticipate challenges, mitigate risks, and ultimately, deliver exceptional products on time and within budget.

Ready to unlock predictable delivery in your organization? Implement the Timeline Rule in your next product planning meeting. Start with one simple question:

What’s the right precision for your next milestone?

Keep the conversation going

Share your experiences, challenges, and success stories of delivering with precision on LinkedIn. Let’s build a community of product leaders who are committed to delivering with precision.

_This post was hand crafted with ❤️ using AI_

“A black box.” That was my candid assessment of Large Language Models (LLMs) for a long time. Despite my background in building neural networks, I viewed Generative AI primarily as a productivity hack—sophisticated autocomplete that was impressive, but too prone to hallucination for mission-critical enterprise applications.

I went into Google’s deep learning boot camp to challenge that skepticism. I wanted to know if we were building on a foundation of hype or a new substrate of computing.

I left with a completely different mental model. The breakthrough wasn’t seeing the models get bigger; it was seeing how we can control them.

The Strategic Pivot: From Generation to Orchestration

The boot camp dismantled the idea that we are at the mercy of the model’s training data. We explored techniques that turn these “black boxes” into transparent, architectural components.

For product leaders, two specific architectures fundamentally change the unit economics of software:

1. RAG: Solving the Customization Dilemma

We explored Retrieval Augmented Generation (RAG), which grounds the model in external, verifiable data.

Implication: RAG solves the “Enterprise Customization” problem at scale. Traditionally, building a workflow tool that adheres to the unique compliance policies of a Fortune 500 client required massive custom development. With RAG, we don’t rewrite code; we simply connect the model to the client’s own policy documents. The model generates compliant workflows dynamically. This shifts the value proposition from “software that works” to “software that adapts.”

2. Reasoning Chains: The Audit Trail for AI

We also harnessed Chain-of-Thought and ReAct prompting techniques. Chain-of-Thought instructs the model to to output intermediate reasoning steps which reduces hallucination. Where as ReAct (Reason and Act) allows the model to decide when to perform actions (like searching or running code) before answering [https://github.com/ysymyth/ReAct].

Implication: This moves AI from a “trust me” system to an auditable partner. In high-stakes decision-making, we cannot accept black-box answers. By forcing the model to “show its work” and execute deliberate actions, we start to create the reliability and interpretability required for the boardroom.

The New Literacy for Product Leadership

This experience crystallized a shift in the product management capability stack.

Ten years ago, the most effective product managers were those who learned SQL. They didn’t wait for data science teams; they mined their own insights to drive decision velocity.

LLM orchestration is the new SQL.

Tomorrow’s product leaders won’t just write specs; they will architect intelligence. Imagine firing up a vector database, embedding thousands of unstructured customer support tickets, and querying: “What are the hidden friction points for our APAC enterprise users?” This capability allows us to move from analyzing metrics (what happened) to analyzing meaning (why it happened) at a scale that was previously impossible.

What does this mean for the exec team?

The question is no longer “Can a machine understand meaning?” The question is “Can your organization harness this meaning to see around corners?”

We are moving from an era of Predictive AI (what is the next word?) to Agentic AI (what is the next move?). Leaders who understand the architecture of these systems—who understand the difference between a raw model and a RAG pipeline based workflow —will have a distinct advantage in risk assessment, scenario planning, and product innovation.

I walked into the boot camp a skeptic. I walked out a builder. The tools to transform our products are here; it is now a matter of intentionally embracing them.

How is your organization moving beyond the “hype” phase of AI adoption?

Want to know a secret to product leadership success? Effective communication. And a key part of that? Mastering the product update. 🚀 It’s your chance to showcase progress, highlight wins, and inspire action. But it’s also an opportunity to demonstrate your leadership and communication skills, which are crucial for career growth. A well-structured product update serves as a critical instrument for fostering alignment, driving collaboration, and securing stakeholder buy-in. 🤝 It provides a clear and concise snapshot of project status, risks, and opportunities, enabling informed decision-making at all levels of the organization. 💡 Furthermore, it showcases the product leader’s ability to communicate effectively, a crucial factor in career advancement.

Early in my career, I learned a valuable lesson about the power of concise communication. A senior executive once remarked that my updates were ‘information overload.’ 🤦‍♂️ That feedback prompted me to rethink my approach, focusing on delivering key insights and actionable information rather than overwhelming stakeholders with unnecessary details. The result was a significant improvement in the effectiveness of my communication and the receptiveness of my audience.

Three Key Principles for Effective Product Updates

1. Outcomes over Output:

ocus on the ‘so what?’ 🤔 Stakeholders are primarily interested in the tangible results of your team’s efforts. Quantify your achievements whenever possible. Instead of stating ‘We held several meetings this week,’ articulate the outcomes of those meetings: ‘We secured agreement from key stakeholders on the revised product roadmap.’ ✅

2. Facts to Beat Fiction:

Data-driven insights are essential for effective communication. 📊 Replace subjective assessments with concrete metrics. Instead of saying ‘User engagement is up,’ provide specific data: ‘User engagement increased by 15% this week, exceeding our target of 10%.’ 📈

3. Clear Action Items:

Eliminate ambiguity by assigning clear ownership and deadlines for all action items. 🎯 A well-defined action item includes a specific task, a named individual, and a firm completion date. For example: ‘John Smith will finalize the design specifications by end of day Friday.’ 📝

By implementing these principles, product leaders can transform their product updates from routine reports into powerful tools for driving product success. ✨ Embrace clarity, conciseness, and a focus on outcomes, and you will see a marked improvement in the effectiveness of your communication and the impact of your leadership.

This post was hand crafted with ❤️ using AI.

Meetings are the lifeblood of people-centered work.

But most of us are not adequately trained in the practices and tools needed to manage meetings. i.e. planning, facilitation, communication, decision making, and conflict management.

Part I covered how to prepare to have effective meetings. With an agenda published you now have a team of humans ready to make some decisions around a topic. So let’s talk about the process of running the meeting.

Open with a Clear Plan

Be on Time, Be prepared, and Start on Time

You’ve invested a lot of time in setting up the meeting time and your teammates are stopping their other tasks to attend your meeting. Start your meetings on time to allow ample time for discussion.

If you’re presenting be sure that you’ve already solved the “technical difficulties” ahead of time.

Use the “Opens” Process

At times, new and pressing items come up between publishing the agenda and meeting. We call these “Opens.” This process surfaces and handles the top of mind issues that might derail your meeting.

First collect the opens. The team goes around the room to see if anyone has any pressing issues. Only capture the item on a whiteboard or in a shared doc. This is not the time to go into detail or discuss them.

Address each Open item

The meeting owner goes through each item in the opens list. If the item is <1 min update and there is time allocated to “Opens” then cover the item.

If the item requires a discussion the group decides where to add that open item into the agenda (ie. which agenda item it should replace) or the group can call another meeting.

Keep the Meeting On Track with Roles

The majority of the meetings become unproductive in this phase. It is important to keep discussions on track and the team moving towards their goals in this phase. There are a few other roles that will help you manage the meeting.

The Timekeeper

First, each agenda item should have a specific amount of time assigned to it.

Choose one person to ensure that you’re staying on the time that you allocated to each agenda time. This person is also responsible for leading the meeting to an on-time conclusion.

The Decision Maker

For each agenda item define the decision-maker and the decision-making process. Major processes people use are: Consensus, Majority or other Voting, Consultative, Directive/Unilateral. This removes confusion for the team by signaling the kind of input the team needs to provide and when the discussion will end.

Source:Decision Making Styles - Next Group

The Scribe

This person captures and communicates the decisions form the meeting. Critically, they are responsible to share next steps AKA Actions Required. Assign someone at the start of each meeting and rotate the role for recurring meetings.

“AR”s have owners and a due date. Someone who is not at the meeting should be able to understand what should happen.

I prefer the following format for simplicity and consistency.

For example: Date Assigned - Item ## - Action Item - Owner - Due Date

The Includer

This is a role that each participant can and should play. This role ensures that all voices get airtime. They intervene to bring in people who have not said anything during the meeting into the conversation. A simple intervention is to notice who is not talking and ask “Well, X, What’s your opinion?”

A second intervention is to ensure there is only “One meeting” at any time. This means there is only one speaker at a time. Don’t start a sidebar or engage in a sidebar when someone else is talking.

The Includer is a key role that makes sure that the loudest voices don’t dominate the room.

Concluding the Meeting

An effective meeting sets the stage for clear follow-up. Leave yourself time to wrap up with clear action items and a call to action.

Readout and Review the ARs

The scribe should read out each action item as captured (not as interpreted.) This is the time for owners of the action items to clarify any information that they need to complete the AR.

After this, it is the owner’s responsibility to disposition the AR before the next meeting or due date.

Publish Minutes

This is a critical step in the Meeting and should be published to all stakeholder within 24hrs.

Many people write minutes as a transcript. I find that it’s generally more efficient to succinct summary of the discussion and decisions- not a play by play. However,if trust is low a transcript might be more appropriate.

Have Fun

Effective meetings speed up the exchange, debate, and discussion of information critical to teamwork. No one wants to spend more time in meetings. With these tips, my hope for you is more effective meetings that drive the results you and your teams are capable of.

If you put in place these or find other tips and tricks for great meetings - I would love to hear from you. Twitter

The Problem

Millions of meetings. Each Day. They’re mostly bad

Each day the US “holds 55 million meetings” according to a recent Freakanomics episode on NPR. Getting together with people should be a happy and playful experience. Yet, mention the word “meeting” and watch the fun leave 55 million faces.

No Meetings: Not an option

The knee jerk response to bad meetings is to eliminate ALL meetings. But reality is that “no meetings” teams end up with all the same meetings with fun names to fill role that the old meetings used to serve. Meetings are the processes, rituals and places we leverage to bring people together to a common purpose. If you aim to work with people and tackle complex, meaningful, work you will need meetings.

Meetings are hard work and we’re not well trained

Effective meetings need a combination of planning, facilitation, communication, decision making, and conflict management skills. There are many consulting careers built on each of these fields. Yet many of us have little exposure or practice in them.

Solution

The Effective Meetings Cheat Sheet

The best meeting culture I’ve experienced was at Intel. Every employee takes a class called Effective Meetings. Effective meetings lays out a clear meeting framework: planning, execution and follow up. These simple tips and effective practices for before, during and after the meeting will level up your meetings.

Part 1: Before the Meeting

1. Do you really need a meeting?

   Don’t have a meeting to “present” output or get facetime. If you’re only communicating status and everyone is going to continue according to the prior plan. Send the status update by email/slack/carrier pigeon as appropriate. Only meet to co-ordinate a change of action.

   Eg: You realize that you need marketing to change their ad campaigns based on this week’s result. It requires rapid and complex coordination across teams. This is worth having a meeting.

2. Know why you’re meeting and the type of meeting you need: Process vs. Mission

   • Process Meetings: Regularly scheduled to coordinate action fixed agenda - eg: 1-1s, Core Team Meetings.
   • Mission Meetings: Ad-hoc meetings to solve a specific problem and/or produce a decision.

3. If you still want the meeting, Publish the Agenda

   The agenda defines the purpose of the meeting. Eg: What decision are you making? What outcome do you want? Publish the agenda when you call the meeting.

   A good agenda includes: the purpose, the items to discussed, time allocated for each item, any roles and decision making process. (Why, What, When , Who, How)

   Example Agenda
   
   Meeting to Decide on Color for the Bike Shed
   
   * Collect Opens Sandeep 2 mins
   * Review Customer Feedback (Paul - CS - 4 mins)
   * Understand Budget Impact of color choice (Peter - Finance 4 mins)
   * Final inputs on  Color (All - 4 mins)
   Decision Maker: Jenny - Consultative
   Minutes: Jack
   

4. If you’re invited to a meeting know that all meetings are optional

• You’re not obligated to attend a meeting because you recieved an invitation. You should attend if you can contribute towards the meeting’s purpose.
• On that note, if there is no a published agenda - there is no purpose and you are not obligated to attend.
• If the agenda doesn’t need your input and you’re are not involved in the decision making process. You can skip the meeting and wait for meeting minutes to be published.

What’s Next?

By taking the time to articulate a clear and crisp Agenda you now have the right people in the room, clarity on what decision needs to be made and the each person’s role in the decision making. Congratulations, you’re already in better shape than the average meeting.

In the next post, I’ll cover effective Time Management, Meeting Roles, Decision Making Process, Minutes and Follow Up in order to conduct and conclude your meetings for greatest impact. Follow me on twitter or Linked In if you want to be notified of the next post.

PS - If you worked at Intel and remember something different or something that I missed - shoot me a note.

Clear and effective communication remains a key challenge in the workplace. A well-written status report breaks down communication barriers and keeps your team aligned.

But first, what’s in it for you?

If effective communication is not reward enough, writing a weekly update also levels up your work game in three ways:

1. Helps you focus: Reflection is a critical part of productivity. Writing a weekly formalizes time to reflect on your progress this week and plans for next week. The deliberate practice acts provides a clear focus for your performance.

2. Makes you a better teammate: You might not have a chance to present at “the big meeting” every week. But, with your weekly gives you visibility across the organization. The highlights showcase where you might be able to help. And, the lowlights maybe areas your team can unblock you. You will be surprised who might have a creative solution or the resources you need to solve the issue.

3. Remember all the things: Finally, the weekly serves as an ongoing journal of your work. At end of the year, you don’t have to rack your brains to remember all the wonderful work you did over the 52 weeks. Search your mailbox, Slack or Dropbox and you have a week by week playback of your entire year.

So how do you write a good weekly?

The overarching goal of the weekly status report is to communicate with a bias for action. Clear and succinct writing the removes any cognitive hurdles for the reader that will cause them to TL/DR.

Here are some tips that I have compiled over the years:

1. Cut the Formatting

A weekly report should focus on clear communications not on an overwrought template. It should be easy for you to put together and for the reader to absorb. Avoid spending any time on formatting, fancy graphics, PDFs or create artifacts for their own sake. A simple text email with bullet points is a very effective way to drive the focus to the communication.

Additionally with plain text you can:

• Copy and paste into other project level or organization level reports.
• Read your weekly across any device or by a device.
• Search for your weekly in emails, slack, dropbox/box/google drive etc.

2. KISS Structure

The report needs to cover three major topics:

• What progress have we made towards the goals and objectives?
• What were the challenges that kept us from making progress this week?
• And what are we going to do about it?

I like to cover this under the simple headings: Highlights, Lowlights, and Plans.

Limit the highlights and lowlights sections to 3 to 5 major items rather than a perpetual laundry list. In the plans section you should provide outline the major tasks/next steps for the next week or two.

Here is a template to get you started.

The KISS Weekly Template

Subject: Weekly for <Your Name> - Week Ending XX,XXX, 2019
Highlights - What were the Top 3 things you accomplished towards your goals this week.
1)
2)
3)
Lowlights - What are the 3 Biggest things that didn't go as expected this week  (If you need resources this is the place to indicate)
1)
2)
3)
Plans - What do you plan to do next week
1)
2)
3)
N)

​

3. Clear, Concise, Comprehensible Content

A common mistake that many people make is to turn status reports into a weekly log of every action or superflous narrative describing each meeting. Your weekly should focus on outcomes and showcase your progress towards goals.

Highlights should communicate progress towards completing your objective and key results. Lowlights enumerate the risks that were introduced or inccured towards goal completion. Finally, Plans idenitify the steps you will take to mitigate the risks and continue progress towards goals. It’s important to provide the reader enough context to understand the issue at hand.

Here are a couple of examples of common issues with weeklies and how to rewrite them to be outcome based.

No Good

• Had a meeting w/ a customer (Ed: Not sure what the meeting was about. Or what the outcome was.)
• Went to daily meetings. (Ed: So what?)
• Looked for a contractor. (Ed: Why do we need this? How did it go? Is it done?)

Good

• Positive Introductory call with Customer X. Customer X wants to hear a specific proposal. Plan to follow up in 3 weeks after finalizing the pitch.
• Many last-minute miscommunications this week. Started new daily sync until the end of the project to tackle issues.
• Final version of the project needs a deliverable but we don’t have the skillset. Started contractor search to deliver project on time in 2 weeks ($5K.)

Following these three simple tips will make your weeklies more powerful, delightful to read, easier to write and make you a rock star communicator.

One final bonus tip: Even if you’re drafting weekly masterpieces, do your audience a favor and keep your weeklies opt-in.

Many thanks to my friend and awesome product leader Lily Liang for reading an early draft of this post

Yes but can you show it to me on a map?

Yelp recently published their annual list of the hottest places for 2017. As I mentioned in 2015 I prefer to see this list of restuarants on a map.

So without further ado here are 2017’s 100 Hottest Places to eat, according to Yelp, presented on a map. For a good measure, I’ve included the 2015 and 2016 ones as well.

Built with the Yelp API and Leaflet.js<– amazing, easy, mobile friendly maps.

Source Code on Github

Part 3: Search Front End

Prerequisites: Node.js; Bower

So far we’ve created the search infrastructure, indexed our photos into our ElasticSearch instance and then used Kibana for intial data exploration.

Now we need a simple front end to collect user input and query the back end and tie it all together by displaying the pictures that have the chosen color.

Most of the AngularJS code needed is boilerplate. Rather than recreate the wheel, our app leverages the angular-seed app template and this guide from SitePoint on building a very basic angular + elasticsearch app.

Both those sites provide detailed walk-throughs so the focus of this write up will be on the getting the query right which is the magic behind the curtain for our search experience.

As always, the full code for this example is available in the photosearch project repository. You can run it with npm start. Be sure to change the code to point to your Elasticsearch instance and update the path to your photo library in the view so that they display correctly.

Finding the Right Query: RGB, HSL, LAB, OMG

Color matching, it turns out, is a non-trivial problem. To summarize succintly, RGB space is perceptually non-uniform so your every day euclidean distance formula doesn’t work to give you color that your percieve to be similar. To overcome this, professionals switch to a more uniform color space and then use a distance formula complicated by additional “weights” that account for differences in perception for a given application of the forumla.

To make life easier, Elasticsearch has a powerful concept called filtered query which allows you to aggregate and score the results using “nearness” functions that make the color matching problem simple. This is demonstrated in this post about search by color for an ecommerce store.

We implement a similar scoring function on HSL values that we extracted using our indexer to get a result of photos that have perceptually similar colors to the color selected by the user.

Finishing Up

With the right query in place we execute the search and… Voila!

In an very short time we’ve created a full featured search engine for your photos. Hope you enjoyed this walk-through on how to create a powerful search engine based on some pretty easily available off-the-shelf components!

If you found this interesting and want to chat, feel free to reach out via Twitter:@_sandeep

Part 2: Adding Photos to the Search Engine

This is the second part of a three part series on how to sort through a large photo collection easily. This part covers Indexing: getting information about the images in your photo library into the search engine that we built last time.

Prerequisites: Completion of Part 1; Node.js; Photos with EXIF Data incl. GPS; node-gyp and Cairo which are used by node-canvas;and exiftool. H/T @mmigdol

We’ll set up the Elasticsearch index to hold data about the photos in our collection, extract the relevant data from the images, and add it to the search engine. Finally, a quick visualization with Kibana cofirms that we ingested the data correctly.

If you’re in a hurry, grab the entire code from the repository, install the dependencies with npm install and kick off the indexing process with node index.js <path to photos>.

Setup a Mapping for Photo Search

While Elasticsearch is pretty smart we can help it better understand the kind of data we’re sending by creating a “mapping”. A mapping tells Elasticsearch which fields contain numbers, dates, dominant colors or geolocations and how to interpret them in its aggregation.

Let’s start by using the elasticsearch node.js library to connect with our elasticsearch instance to set up a mapping for common EXIF fields and the colors that we want to store.

BTW - My script is an extension of this neat little script on ingesting photo metadata into ES.

Extract the EXIF and Color Data from Photos

Digital cameras store a set of metadata called EXIF data within each photo. If you’re using your phone or a fancy new camera they will also embed the GPS coordinates of each picture within the EXIF data.

We are going to use the Node.js directory walker to find and process each file in the source directory and extract data from each image. We’re going to use two fucntions to extract data for our little search engine: 1) EXIF Data and 2) Dominant color data.

The second extraction in the code above calls a getPalette function which extracts the 5 most dominant colors in each image so that we can search by image color. Originally, I naively used RGB color extraction but to get perceptually similar colors (ie. ones you precieve to be similar) without a lot of gymnastics we need to use a perceptually uniform color space such as HSL. The palette module converts between color spaces out of the box.

If in the future we want to extend our search engine by building an image or mood classifier to use with our images, we simply need to add another extractor function.

Add Data to Elasticsearch - Bulk Upsert

Since we have multiple asynchronous operations to extract data, we want to make sure that our writes to the Search Index are UPDATES and not CREATES.

Elasticsearch has a nice UPSERT “method” which will update if the Index contains a document with that id or otherwise creates the document.

The sendToElastisearch function collects the output of the extractor functions and sends a bulk update to the ElasticSearch.

Mapping Photo locations with Kibana

Kibana gives us a quick visual way for us to explore the search engine data. Setup Kibana by pointing your browser to your Kibana Installation:

• Select Settings
• Uncheck Index contains time-based events select box
• Set the index name to photos
• Click the Create button

Now, to build a heat map:

• Select the “Visualize” tab
• Select “Tile Map” and”From a New Search.”
• Once the Map loads, select “GeoHash” from the “Select Buckets Type”. Kibana is smart enough to realize that there is only one Geohash and selects it for you as the source in field.
• From “Options” select “Heatmap” as the “Map Type” and adjust the sliders to your taste.
• Press the green arrow.

We instantly have a nice mapping of every picture in our index and can easily create searches and visualizations based on any of the EXIF data that we’ve ingested into Elasticsearch.

In part 3 we will finish up this project by setting up a simple front end with a light footprint to easily find pictures with a certain color.

Motivation

If you’re anything like me, which I suspect you are, over the years you’ve probably amassed a large collection of photos. With over 73,257 photos across 707 directories representing 15 years of digital photography, finding a particular photo is a challenge.

While most photos have some combination of date and time or even GPS information, most people don’t spend their lives memorizing times and dates of photos. We need to augment the “metadata” with something more human.

Wouldn’t it be great if we could search for photos by features that people remember? Features, such as the color of the sunset or the blue of the sky in the picture they’re looking for?

A Three Part Solution

Our solution needs three parts: First, we need robust search engine infrastructure. Second, the photos need to be cataloged and added to search engine’s index. And finally, we need a simple way for users to search through tens of thousands of images by metadata, colors or any other feature we choose to add to the index.

Part 1: Search Infrastructure in a Box

Prerequisites: You will a need working Docker installation.

One of the most flexible search tools available today is Elasticsearch. This provides powerful indexing, query and aggregation functions out of the box. Best of all, you can access all of this via a http API.

In addition to Elasticsearch, we’ll also bring up Kibana, a visualization tool to explore our data.

Loading up with Docker

Instead of setting up these tools individually let’s use Docker to bring up the official containerized distributions of each of these components in seconds (with functional defaults.) Docker’s new docker-compose command starts and links multiple containers together based on the configuration in a docker-compose.yml file.

The configuration file below tells Docker to pull down the official images for Elasticsearch and Kibana from Docker Hub, set up a custom configuration file and connect the Kibana container to the ElasticSearch container.

You can clone this repo to get started on the photo search project.

Start Your (search) Engines

From the project directory simply issue docker-compose up and watch docker create your search engine. (You can use docker-machine on OSX if you need a docker host.)

If you go to http://<dockerhost_ip>:9200 a JSON response with the name of the cluster and Elasticsearch version information verifies that Elasticsearch is up and running.

Verify that Kibana is running and that it can connect to Elasticsearch at http://<dockerhost_ip>:5601

Next Steps

We’ve created a powerful foundation for search that can both handle large amounts of data and be scaled out to service a high volume of requests.

In Part 2, we will extract data from our images and add them to our search index.

PS - You can also check out Part 3, for the AngularJS front end for our search engine.