Alvin OmozokpiaAlvin OmozokpiaAI Product Manager

Blog

Notes from building, studying, and leading products.

Everything here lands in one of five piles. Strategy is what to build and why. Shipping is getting AI out the door with evidence. Trust is the safety work that arrives late. Decision Docs are the real documents behind real calls. Teardowns take one AI mechanic apart at a time. (Yes, Google, I see you 👀)

A small winners podium on a workbench is turned to reveal its hollow scaffolded back while a desk lamp lights its confident front face, teal threads running to a stack of folders.

Teardowns

Teardown: The Leaderboard Is a Product. You Are Not the User.A 16-month study of Chatbot Arena found private variant testing, selective score disclosure, and unequal data access shaping the rankings teams treat as ground truth. The teardown lesson: leaderboards gate attention, and your release pipeline has to gate everything else.A product builder in a mustard sweater gently slides one paper card back across his desk while accepting another into a wooden tray, a tall stack of requests waiting beside the lamp in his warm study.

Decision Docs

Decision Doc: A PRD Where the Core Feature Is Saying NoThe sanitized PRD for teaching a compliance-bound research assistant to refuse low-confidence requests, which cut unsupported claims from 14 percent to 1.8 percent at the cost of a 9 percent refusal rate.A product builder holds a sheet of paper to the lamplight to compare its faint curve with a second sheet on his desk, the two linked by a thin teal thread in a warm study.

Decision Docs

Decision Doc: The Experiment Brief That Replaced Our Offline EvalsThe actual experiment brief I wrote when our offline evaluation scores turned out to have near-zero correlation with production outcomes, and the shadow-mode replay design that replaced them.A product builder rests a small charcoal block on a winding paper path between two open folders on his desk, pausing the route to a little wooden gate under warm lamplight.

Decision Docs

Decision Doc: The Memo That Stopped a CTO-Sponsored AI InitiativeThe actual hold memo I wrote after a three-week feasibility test found a 71 percent accuracy ceiling against a 95 percent requirement, and how the team was redirected without losing the sponsor.Eight upright paper answer cards stand confidently on a reading desk while a brass magnifying glass reveals the blank back of one, a teal thread tracing to an open book under lamplight.

Teardowns

Teardown: AI Search Ships Answers. The Missing Feature Is Doubt.Eight AI search engines answered wrong more than 60 percent of the time on a basic citation task, and almost never declined to answer. The teardown lesson for product teams is that abstention is a feature you have to build, budget, and own.A product builder links an observation card through an interpretation lens to a branching assumption seed.

Strategy

A Product Note Is Not Evidence YetA note becomes useful when the observation, interpretation, and affected assumption stay connected. Living Fieldbook is a prototype for making that chain visible.A product builder reviews memory cards linked by teal threads to an open journal, with an eraser nearby and one trail fading toward a shadowed shelf.

Trust

When Helpful Memory Becomes SurveillanceAI memory earns its place when people can see what was retained, why it matters, where it is used, and how to change or remove it.A product builder pulls one consequential paper case from a crowded queue and follows its evidence trail toward three human review paths.

Trust

The Exception Queue Is the ProductHuman review is not a safety feature by itself. The queue must show why a case arrived, what could go wrong, who can act, and what the system should learn from the decision.A product builder marks a clear boundary between a user-controlled career workspace and an employer decision gate.

Trust

The Compliance Boundary PMs Must Draw Before RoadmappingA career assistant that helps a person tell their story and a system that ranks people for an employer may use similar technology, but they create different product, evidence, and regulatory responsibilities.A product manager at a workshop gate examining several AI outputs under a warm lamp before allowing one to continue

Shipping

The AI Feature Ship GateAn eval should behave like an executable product requirement: it defines the expected outcome, the unacceptable failures, and the evidence required to launch.A product manager tracing a path from many glowing usage dots to one completed outcome, with cost and risk weights beside the path

Shipping

Adoption Is Not an AI Value MetricAdoption shows that people used an AI feature. This metric tree tests whether the feature completed useful work at an acceptable cost and risk.A product builder balances paper cost tokens against an answer card, clock, and correction pencil while one teal path continues from several options.

Shipping

The Cost of a Correct AnswerAI economics should be measured through correct outcomes, recovery burden, latency, and reliability, not a token price or unit-cost headline in isolation.A product builder controls the final glass gate in a five-stage path while a teal AI light handles lower-risk tasks.

Trust

The Agent Autonomy BudgetA practical ladder for deciding what an AI agent may recommend, draft, or do, before convenience turns into an unacceptable product risk.A product builder uses a stop lever and replay loop at a calm incident desk while one red signal marks an AI failure.

Trust

The Incident-Ready AI Product ManagerWhen an AI feature fails in production, the PM needs a severity model, containment decision, user remedy, replay path, and a way to turn the failure into a stronger ship gate.A product builder inspects a hidden red thread moving through a document, an AI tool chain, and an outbound gate.

Trust

Red-Team the Product, Not Just the ModelA model can refuse a malicious prompt in a lab and still sit inside a product that leaks data through tools, permissions, identity, or recovery gaps.A product manager inspecting paper and light data streams at symbol-only checkpoints before they enter a small AI workshop

Shipping

Your AI Product Is Only as Ready as Its DataChoose the use case before the model. This Data Readiness Contract makes purpose, provenance, permissions, quality, freshness, evaluation, retention, and ownership explicit.A product builder raises one hand beside a glowing AI path while source pages, a forked route, and a corrected card represent evidence, uncertainty, stopping, and recovery.

Trust

Designing AI That Knows When to StopA trustworthy AI experience does more than display confidence. It helps people inspect evidence, recognize uncertainty, stop safely, and recover when the system is wrong.A thoughtful product builder at a laptop beside a product flow, a user profile, a chart, and a light-bulb sketch.

Strategy

A Product Why Must Change a DecisionPurpose is useful only when it changes scope, sequencing, evidence, or a stop decision. This brief turns an inspiring why into an operating constraint a product team can actually use.A product builder working at night with a laptop beside achievement, calendar, and progress panels for Bragora.

Shipping

Building Bragora Around User ControlBragora began with a simple product problem: useful career evidence disappears when people need it most. Building the product taught me to treat capture, reuse, and AI control as one connected workflow.A product builder speaking into a desk microphone while an AI assistant responds on a laptop beside workflow diagrams.

Shipping

The Art of the Ask: Prompting as a Product DisciplineThe best prompts are not clever incantations. They are small product specifications that define the task, context, constraints, evidence, and acceptance test before an AI system begins.A warm editorial scene with a phone showing a language-learning path, a green owl figure, a notebook, and celebratory confetti.

Strategy

Duolingo Designed for the Return, Not Just the StreakDuolingo's product lesson is deeper than gamification. Its habit system reduces the cost of starting, makes continuity visible, creates recovery paths, and keeps the learning outcome close enough to matter.A product builder examining a growth chart while customer and product blocks move through a warm, grainy workspace.

Strategy

Product-Led Growth Starts After Sign-UpA free tier can generate accounts without creating a growth engine. Product-led growth works only when users reach value, return for it, and encounter an expansion path that fits the job they are already doing.A product builder working with a glowing AI assistant beside product flows, research notes, and an outcome chart.

Strategy

AI Can Do PM Tasks Without Owning the Product DecisionThe useful question is not whether AI can write a PRD or analyze feedback. It is which product tasks can be delegated, what evidence the system needs, and who remains accountable for the decision.A product builder reviewing an AI-generated prototype on a laptop surrounded by research notes and product charts.

Shipping

Faster Prototypes Need Better EvidenceGenerative AI can make a prototype cheap enough to build before a meeting. That speed is useful only when the team is explicit about which assumption the artifact tests and what evidence would change the roadmap.