Writing
Essays on building AI products that ship — reliability and evaluation, agentic systems, and the product craft in between.
Is it production-ready? The four metrics that decide whether your AI agent ships
Most enterprise GenAI pilots don't die because the model is weak. They die because nobody can prove the agent is good enough to trust. Here is the evaluation layer that turns 'it feels good' into 'it's ready'.
Read the post →Human-in-the-loop is a product decision, not a safety checkbox
Teams bolt a human onto AI to feel safe, then wonder why nothing gets faster. Where the human sits is one of the most important product calls you'll make — here's how to make it deliberately.
Build, buy, or orchestrate: a PM's framework for enterprise AI agents
The old build-vs-buy question has a third answer now, and it's usually the right one. A practical framework for deciding what to build, what to buy, and what to simply orchestrate.
From one-off to catalog: turning AI builds into reusable agents
The first AI agent is a project. The fiftieth is a platform — but only if you design for reuse early. How to stop rebuilding the same agent for every customer.
The PRD isn't dead — but writing specs for non-deterministic AI is different
You can't spec a probabilistic system the way you spec a button. The PRD doesn't disappear for AI products — it changes shape, from listing behaviors to defining acceptable distributions of behavior.
From spec-writer to builder: what shipping StockSnap taught me
I usually spec AI products for engineers to build. For once I built one end to end — a photo-to-inventory app for kirana stores. Shipping it myself quietly changed how I do my actual job.
In regulated industries, compliance is a feature — not a tax
Engineers tend to treat compliance as friction to minimize. In healthcare and banking, the AI products that win treat it as the product — the thing customers are actually buying.