AI systems that hold up in production.

The metric was measuring the wrong thing

The reported distress rate was climbing, 21 percent to 27 percent year over year, and it read like a market signal. It was 187 spurious labels from companies that had simply not filed their accounts. I traced it to the label definition rather than to the model, and added a reportability gate that holds the rate inside a point.

92.155% on current-year distress flagging ✦ SHAP attached to every prediction

₹95,000 a month for bad answers

A facilities management platform was paying a vendor copilot to answer questions about their own data badly. I modelled the unit economics layer by layer before writing code: the cheapest model for intent routing, semantic table selection so only relevant schemas reach the LLM, three tiers of cache.

₹95,000 to ₹35,000 a month ✦ post-deployment spend tracked the model

developers across nine coding tools

A tool-agnostic memory layer that learns a developer's habits from real edits and writes them into whichever coding agent they are using. One thousand and sixty-seven tests across sixty-six files, twelve of them security suites, after a hardening sprint that closed symlink races, Unicode sanitiser bypass and indirect prompt injection.

eleven stars ✦ security audit public in the repo

It answered fluently and knew nothing. Retrieval was querying an empty collection, so the model fell back on itself. Silently.

fix ✦ check the collection at startup, disable retrieval when empty, say so in the log

I model the unit economics layer by layer before writing code. Cheapest model for routing, semantic table selection, tiered cache.

then ✦ the projection gets checked against real spend

A calling product was hanging up mid-call. I built a diagnostic instead of an architecture.

found ✦ all nine recordings were 8 kHz mono, so barge-in cannot be measured