Projects

Detects civic issues via CCTV/dashcam, auto-generates complaint tickets, verifies repairs, and retrains itself.

• Model & Inference: Trained YOLOv8 on 4,783 images with 30% engineered negatives and FN-bucketing for recall diagnosis; achieved 20ms ONNX inference on a FastAPI microservice deployed via GCP Cloud Run.
• Auto-Ticketing & Reliability: Engineered a burst-frame extraction pipeline feeding a 4-tier reliability engine and DIGIPIN geospatial deduplication (4m² grid), processing 256 end-to-end complaints in 0.36s per ticket.
• Verification & Active Learning: Enforced automated post-repair rescans to verify fixes before ticket closure; bucketed field data into labeled classes for continuous baseline-gated retraining.
• LLM Routing (Seva): Integrated Gemini 2.5 Flash using a 4-step Chain-of-Thought to autonomously route complaints across 42 civic categories, auto-routing 125/256 tickets with zero manual input.

5-agent pipeline: legal case in — research, strategy, drafted documents, and reasoned explainability out.

• Orchestration & Intake: Built a stateful 5-stage LangGraph pipeline (Intake → Research → Strategy → Drafting → Explainability) with OCR-aware processing and SHA-256 caching to mitigate LLM hallucinations.
• RAG & Determinism: Indexed 4,582 chunks (7 legal acts) in Qdrant via Sentence Transformers; engineered a diagnostic harness to catch retrieval drift and OCR instability across N-run tests.
• Reliability & Graph: Implemented a triple-engine fallback (Groq → Gemini → offline rules) for zero-downtime generation; built a Neo4j legal knowledge graph (1,410 nodes, 1,837 relationships) for explainable document drafting.
• Realtime Infrastructure: Developed an event-driven custom Redis Pub/Sub, WebSockets, and Celery backend for live citizen-lawyer negotiation chat, bypassing third-party APIs.