Just shipped something I'm proud of: a CLI tool that runs 4 concurrent AI review passes on your git diffs before you push.

It catches bugs I would have missed. Not theoretical bugs — actual logic errors in production code.

The stack: Python, Claude API with structured outputs, ChromaDB for context.

What I learned building it: structured outputs from LLMs are a game-changer. Instead of parsing free-form text, you get typed, validated responses every time.

If you're building AI tools, stop treating LLM output as strings. Treat it as data.

Link in comments. It's called preflight-ai and it's on PyPI.

#AI #DeveloperTools #Python
