I build thething first.Then I learn itinside out.

An environment for catching AI agents lying about their own work. An overseer model reads another agent's tool-use trace, decides if it was honest, lazy, reward-hacking or deceptive, and has to point at the exact step that gave it away. Big models could name the crime and never the moment. So I paid a small one only when it named both, and trained it until it did.

what I didn't know going in — I had never trained a model with GRPO, never touched TRL, and had a rough idea of what LoRA even was. Three days later there was a chart, a live server and a test suite.

Someone had two hundred PDFs and no patience. This reads them, falls back to OCR when the scan is bad, tags and classifies each one, and then lets you just ask questions about the pile in plain English. Multiple documents at once, and it keeps its head across a long conversation.

what I didn't know going in — Chunking and context windows. My first version confidently answered from the wrong document, which is a fun way to learn what retrieval actually has to do.

I got tired of writing the same joins to answer small questions. Ask it in English, it plans the query, runs it against the schema and hands back the answer with the SQL it used — so you can check its homework instead of trusting it.

what I didn't know going in — How much of a good agent is actually guardrails. The model part took an evening. Making it safe to point at a real database took the rest of the week.

Started as a favour for someone and ended up shipped like a client job. The database sleeps when nobody's using it, so the interface pretends the write already worked and quietly catches up in the background. Nobody who uses it has ever noticed the lag.

what I didn't know going in — Optimistic UI, and how much of good product feel is just hiding honest latency from people who don't care about your infrastructure.

500+ teams entered, 25 or so reached the grand final. Ours was a super-resolution model for ISRO thermal imagery, judged in person over 36 hours.

Not a build — just a reminder that I read the documentation for fun and it occasionally pays for itself.