GeoMinder

A multi-tool agentic RAG system for mining consultants who lose 30–40% of their time searching a ~100k-document technical corpus (EIAs, regulations, design memos). The agent answers qualitative questions over documents, quantitative ones over environmental measurements, and chains both to answer “are we within regulation?”. Evolved from a TRL 6 Streamlit demo to a TRL 8 production architecture on GCP.

Highlights

• Event-driven ingestion. A Cloud Run Job triggered by Eventarc on every GCS upload keeps the Vertex AI RAG corpus in sync, with manual full re-sync and scheduled defensive reindex modes, all emitting structured JSON logs.
• Guardrailed analytics tool. Instead of free-form text-to-SQL, a parameterized BigQuery tool with pre-approved query shapes, bound parameters (zero injection), and a dry-run cost check that aborts expensive queries before they spend.
• Evaluation as a CI gate. Golden set + deterministic metrics + LLM-as-judge groundedness (Gemini 2.5 Pro judging 2.5 Flash to avoid self-grading), blocking merges below a 0.85 pass rate.