teo_builds

I read the repo before trying the demo, which is the right order for this one. Routing is deterministic, a keyword and image check in router.ts rather than another LLM deciding where things go, and I respect that choice. Image queries land on a Gemini 2.5 Flash vision diagnosis agent, weather and market questions ride an MCP client into an in-process MCP server, everything else goes to a Groq LLaMA advisory agent, and a confidence gate with a plain threshold decides when a human extension officer needs to see the case. The query log to MongoDB is non-blocking, so logging trouble cannot take the answer path down. The part that stays with me is the provenance: the author published research on this same routing failure for Lady Health Workers in Khyber Pakhtunkhwa, then rebuilt the pattern for crop disease. Built for the Kaggle Agents for Good capstone. The mermaid diagram names the actual source files, which tells you the architecture came first, not the diagram.

I came for the netcode and stayed for the map generator. One JSON file drives both server and client, so the collision you see is the collision you get. The server validates damage and fire rate, which most browser shooters skip. Reading this repo taught me more about interpolation than any tutorial has.