InfraLens Georgia

Privacy-preserving infrastructure intelligence for Georgian municipal incidents. Submit a free-text report, get a structured, cited, bilingual brief — without sending anything to a hosted LLM.

Municipal infrastructure reports arrive as unstructured Georgian or English text. Decision-makers need consistent classification, severity, and procurement evidence — but most AI tooling either ships the data offshore or hallucinates without grounding.

• 01Designed a LangGraph agent pipeline with five sequential nodes — ExtractEntities, ClassifyIncident, RetrieveEvidence, GenerateBrief, VerifyGrounding — each traced and confidence-scored.
• 02Implemented hybrid RAG: pgvector semantic search fused with PostgreSQL full-text search via Reciprocal Rank Fusion (RRF) for higher recall.
• 03Reports below 0.6 confidence are automatically flagged for human review.
• 04Runs entirely local: Ollama with qwen3:8b for generation and bge-m3 for embeddings — no hosted LLM APIs, no paid cloud.
• 05Containerised across 6 services with Docker Compose; FastAPI backend, Next.js frontend, Postgres + pgvector, Redis/RQ worker, Ollama.

• Hybrid retrieval (vector + BM25 + RRF fusion)
• Citation grounding with confidence scoring
• Bilingual Georgian / English briefs
• Local-only inference (Ollama)
• LangGraph traced agent pipeline