Field notes on building AI products fast: the two-model AI engineering team, shipping with AI agents, voice AI for India, and keeping AI-written code small.
Claude Code and Codex build different parts of a task in parallel, then review each other's code. Here is the loop, the rules, and where I step in.
Karkaana is Tinder for factories in India. How it went from a reverse-engineered US platform to 94,000+ indexed factories and 86 commits in about 17 days.
Why Pip? picks a trending topic, writes the script, draws, voices, edits and uploads a Short every day. A walk through the pipeline, station by station.
Zaya's voice agents handle bookings and negotiation in 12+ Indian languages, in real time. Notes on what it takes for a voice agent to feel instant.
One AI-built project grew past 16,000 lines. Since then every task gets a line budget, a failing test before any bug fix, and a cut-down review.
Next step
Seven days from framing to delivery. A two-model AI engineering team does the building, and I sign off on every risky call.