I work on the plant floor and now build AI agents on nights and weekends. The tools change, but the approach stays the same — trace the failure, plan the next move, and keep eyes on what the system is actually doing.
LangGraph, RAG, evals, and output guards. Agents that fail loudly, log honestly, and stay debuggable.
Supply Chain Multi-Agent SystemStart at the failed output and walk back through every condition until the cause is in hand. Years of doing this on real PLCs taught me to make the LLM narrate — not reason about ladder semantics from scratch.
Break the work into shippable milestones with typed state and clear gates. Real-LLM integration tests run in CI, so a regression shows up before merge instead of after.
Output guards with deterministic rules first and an LLM judge second. Self-hosted Langfuse for traces. Nothing ships as a black box.
Real code under CI, all open source. Agent platforms, mobile apps, systems tools.
Three specialist agents coordinating through a typed LangGraph state machine
AI-powered Allen-Bradley PLC ladder logic troubleshooter
From-scratch Rust BitTorrent engine targeting libtorrent-rasterbar parity
Offline-first Android debt tracker, published on Google Play
Single-binary Rust service bridging Telegram to Claude Code sessions