Offline retrieval-augmented AI for transit maintenance
PythonFAISSLocal LLMsStreamlitBART
Overview
Built at BART, this fully offline retrieval-augmented assistant indexes five years of work orders and technical manuals. Technicians describe a problem in plain language and get grounded, source-backed fixes from prior work, with no sensitive operational data ever leaving the building.
Approach
01Ingested five years of work orders and technical manuals into a chunked, searchable corpus.
02Generated embeddings and built a FAISS similarity index for fast semantic retrieval.
03Ran local open-source LLMs for generation. The entire stack runs on-premises, fully offline.
04Built the Streamlit interface and validated recommendation quality directly with maintenance staff.
Results
Technicians receive source-backed historical fixes instead of searching through manuals or asking engineers.
Used cosine similarity to surface similar past work orders as technicians type in a new issue.
The prototype was designed to run fully offline, so no data is sent over APIs to external language models.