I built a local Postgres triage co-pilot because HIPAA says I can't paste plans into ChatGPT or Claude

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

I am a Staff DBA at a top public healthcare company. I look after MS SQL Server and PostgreSQL fleets, and every one of our database servers is on-prem and air-gapped. We do not have the option to use a frontier LLM for production work, because of HIPAA. Query plans, table names, filter literals, none of it is allowed to leave our network.

That means when one of our Postgres instances starts misbehaving at 3 AM, I cannot drop an EXPLAIN plan into frontier models like ChatGPT or Claude and ask what changed. I read the JSON by hand, the way I did in 2015. Most of my colleagues do the same.

I built Plansmith for myself. You feed it two EXPLAIN (ANALYZE, FORMAT JSON) outputs, a baseline from when the query was healthy and the incident plan from now, and it gives you a triage runbook in plain English: what changed in the plan, ranked root cause hypotheses, things you can try in the next five minutes, and the durable fix. The whole pipeline runs locally, including the model. The only outbound network call Plansmith makes is to 127.0.0.1:11434, which is Ollama on the same laptop.