01 / The Challenge
The Problem & Context
India's court system carries a backlog of 45M+ pending cases, and the data that could explain why cases stall — delay patterns, humanitarian urgency, court-level bottlenecks — is scattered across government portals with no unified, queryable view. NITI.AI builds that view and layers judicial-grade AI reasoning on top of it.
02 / Technical Approach
Architecture & Implementation
A scheduled scraping layer pulls live data from 5 government portals every 6–24 hours, feeding a 6-node agentic pipeline: Ingestor → Delay Analyser → Humanitarian Flagger → Scorer → LLM Recommender → RAG Indexer, streamed to clients via SSE for real-time pipeline visibility.
11 FastAPI routers backed by async MongoDB (Motor + Beanie ODM), with JWT auth hardened by TOTP two-factor authentication for sensitive case-data access.
A RandomForest model predicts case delay likelihood from structured case metadata, while a FAISS/ChromaDB-backed RAG layer answers natural-language questions over legal documents.
8 Next.js dashboards (Recharts + Zustand for state) built with an NJDG-inspired information architecture, so the tool feels native to people already familiar with India’s National Judicial Data Grid.
Deployed as a Docker Compose stack, tuned for sub-200ms p95 API latency and sub-1s Groq inference on the LLM-recommendation step.
04 / Technology
Technologies & Tools Used
05 / Impact
Results & Metrics
Key Accomplishments
- Unified view over a 45M+ case backlog that previously existed only as fragmented portal data.
- Sub-200ms p95 API response and sub-1s LLM inference — fast enough for interactive dashboard use, not just batch reporting.
- A 6-node agentic pipeline that turns raw scraped case data into scored, flagged, and explainable recommendations end to end.