Agentic Case-Law Intelligence
Structured intelligence from Indian case law, built as a nine-agent system on GenW.AI and Qwen3.
Impact
2nd / 22,000+
Deloitte Hacksplosion 2026, national
9 agents
Under a supervisor + master agent
Source-verified
Confidence-scored, schema-validated, HITL
Technology
Key decisions
- 01
Nine specialised agents, not one prompt
Each part of a judgment gets its own agent (parties, facts, statutes, reasoning), so extraction is accurate per section and the agents run in parallel. A supervisor and a master agent coordinate them instead of one model trying to do everything.
- 02
A responsible-AI layer before anything is trusted
Every field is verified against the source document, confidence-scored, and schema-validated. Low-confidence output goes to a human review queue instead of being shown as fact, which matters when a wrong citation can mislead a lawyer.
- 03
Built on GenW.AI with Qwen3
GenW.AI's Agent Builder and App Maker let me ship the full agent pipeline and the interface fast, with Qwen3 as the model and PostgreSQL as the structured store.
- 04
A product, not just an extractor
The structured data drives a real interface: a dashboard, case viewer, case comparison, chat over a case, and a human review queue. The output is something a lawyer can actually work in.
Architecture
- 01
Intake Agent
Parses and segments each Indian court judgment
- 02
Master + Supervisor Agents
Orchestrate the run and route extraction
- 03
Six specialised agents
Run in parallel
- Metadata
- Facts
- Petitioner
- Respondent
- Statute
- Reasoning
- 04
Validation & responsible AI
Source check · confidence scoring · schema · human-in-the-loop
- 05
PostgreSQL
Structured store for validated output
- 06
Frontend — GenW.AI App Maker
- Dashboard
- Case Viewer
- Case Comparison
- Chat with AI
- Human Review
Scroll the diagram horizontally to explore →
The problem
Indian courts publish judgments as long, unstructured PDFs. Pulling out the parts that matter (the parties, the facts, the statutes, the court's reasoning) takes hours, and a single misread citation can change an argument. Doing it reliably across thousands of cases needs more than one LLM call.
How it works
I built it as a multi-agent system on GenW.AI's Agent Builder, running Qwen3. An intake agent parses and segments each judgment, then a master agent and a supervisor agent route the work to six specialised agents in parallel: metadata, facts, petitioner, respondent, statute, and reasoning. Before anything is stored, every field passes a responsible-AI layer that verifies the source against the document, scores confidence, and validates the schema, sending low-confidence fields to a human review queue. Validated results go into PostgreSQL and surface through a frontend built with GenW.AI App Maker: a dashboard, a case viewer, case comparison, chat over the case data, and the review queue.