- Built a tool-calling LLM agent on the live BKK API with hybrid RAG (pgvector + full-text, RRF).
- Evaluated 7 models on a held-out set with two LLM judges; 26/26 correct tool calls at $0.001 per query.
- Built a GTFS-RT pipeline processing 800k events a day into a nightly punctuality report.
– present