A B.Tech Major Project (Jan–May 2026) School of Computer Science, UPES Dehradun
Nyāya (न्याय) is a multi-agent AI legal assistant tailored for the Indian legal system. It uses an orchestrator model that dispatches incoming briefs to five specialist sub-agents — Client Intake, Legal Research, Case Strategy, Legal Drafting, and Compliance & Contract Review. The system cites Indian statutes (BNS 2023, BNSS 2023, Constitution, Indian Contract Act, IT Act, DPDP Act 2023) and produces court-ready outputs such as FIR drafts, legal notices, plaints, writ petitions, and compliance memos.
- Five role-specialised sub-agents orchestrated by a router model (built on Mastra)
- Indian-jurisdiction-first: BNS, BNSS, Constitution, Contract Act, IT Act, DPDP, Consumer Protection
- Retrieval-augmented generation over a curated corpus of statutes and compliance rules
- Per-user accounts with bcrypt-hashed passwords, cookie-based sessions
- Persistent chat history (SQLite) — every brief is saved per user
- Server-sent-events streaming with a live agent-flow timeline in the UI
- Markdown rendering of judgments, drafts, and clause-by-clause review
- Offline evaluation harness with confusion-matrix-based scoring
Legal services in India are gated by cost, jargon, and delay. While LLMs can produce useful first-pass drafts, a single prompt-and-respond model conflates the very different skills of (a) extracting facts, (b) finding the right statute, (c) recommending strategy, and (d) producing court-format output. Nyāya decomposes this workflow into five specialist agents and routes every query through an orchestrator, giving auditable, agent-attributed responses that cite Indian statutes correctly and refuse to dispense legal advice without a disclaimer.
| Metric | Single-prompt baseline | Multi-agent (Nyāya) | Improvement |
|---|---|---|---|
| Routing accuracy | 64.2% | 91.7% | +27.5% |
| Tool-call accuracy | — | 92.5% | — |
| Citation accuracy (n=60) | 71.8% | 88.3% | +16.5% |
| Hallucination rate | 12.4% | 4.2% | −8.2% |
| Macro F1 (sub-agent routing) | — | 0.92 | — |
| Disclaimer compliance | 18% | 100% | +82% |
Per-agent classification (test set: 240 cases):
| Sub-agent | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| Client Intake | 0.92 | 0.89 | 0.90 | 36 |
| Legal Research | 0.94 | 0.91 | 0.92 | 58 |
| Case Strategy | 0.88 | 0.85 | 0.86 | 42 |
| Legal Drafting | 0.96 | 0.93 | 0.94 | 54 |
| Compliance Review | 0.93 | 0.95 | 0.94 | 50 |
| Macro avg | 0.93 | 0.91 | 0.92 | 240 |
Latency: mean 18.4 s, p95 47.1 s on commodity hardware.
The pipeline consists of:
- Frontend — minimal HTML/CSS/JS chat UI with a live agent-flow timeline
- Backend (FastAPI) — auth, session management, SQLite persistence, SSE streaming
- Orchestrator (Mastra) — main routing model that decides which specialist to invoke
- Sub-agents — five role-specialised agents (Intake, Research, Strategy, Drafting, Compliance)
- Knowledge Retrieval (in-process MCP) —
retrieve_legal_docsandretrieve_compliance_rulestools over curated JSON corpora (BNS, BNSS, Constitution, DPDP, IT Rules, etc.) - Evaluation harness — offline scorer that runs labelled cases and reports precision/recall/F1 per sub-agent
📸 See ARCHITECTURE.md for the full Mermaid flow + sequence diagrams.
- Orchestrator framework: Mastra
- Backend: Python · FastAPI · uvicorn · SSE
- Persistence: SQLite (users, chats, messages)
- Auth: bcrypt + itsdangerous (signed cookie sessions)
- Frontend: Vanilla HTML/CSS/JS · marked.js · DOMPurify · Cormorant Garamond + Inter
- Knowledge Retrieval: in-process MCP tools over JSON corpora
- Container: Docker / docker-compose
- Routing accuracy — Was the correct sub-agent invoked for the query type?
- Tool-call accuracy — Was the right retrieval tool used when one was expected?
- Precision / Recall / F1 — Per sub-agent classification, plus macro and weighted averages
- Citation accuracy — Manual sample of statute citations checked against ground truth
- Hallucination rate — Fraction of fabricated section numbers or case names
- Disclaimer compliance — Every response ends with the required disclaimer
📸 Confusion matrix and full breakdown in evals/results.md.
- Native Indian-language interface (Hindi, Bengali, Marathi, Tamil) — voice + text
- Live retrieval over Indian case law databases (IndianKanoon, SCC Online, Manupatra)
- Integration with court e-filing portals (eCourts, NJDG)
- Vakalatnama generation and digital signature workflow
- Citation verification layer to drive the hallucination rate below 1%
- Mobile app for advocates with offline draft review
For any inquiries or feedback, please contact:
Name: Deepanshu Miglani Education: B.Tech CSE(AIML), UPES, Dehradun Email: deepanshumiglani0408@gmail.com / Deepanshu.106264@stu.upes.ac.in GitHub: deepanshum0408
Name: Divi Saxena Education: B.Tech CSE(AIML), UPES, Dehradun Email: divisaxena04@gmail.com / Divi.107784@stu.upes.ac.in GitHub: Divi-Saxena
Name: Ayesha Varshney Education: B.Tech CSE(AIML), UPES, Dehradun Email: ayeshavarshney245@gmail.com *
Name: Deepali Rana Education: B.Tech CSE(AIML), UPES, Dehradun Email: ranadeepali45@gmail.com
Dr. Sahinur Rahman Laskar Assistant Professor (Senior Scale) School of Computer Science, UPES, Dehradun, India Email: sahinurlaskar.nits@gmail.com / sahinur.laskar@ddn.upes.ac.in
@misc{nyayaai2026, title={NyayaAI: An AI-Powered Legal Assistant Using Multi-Agent Architecture and Retrieval-Augmented Generation}, author={Deepanshu Miglani and Divi Saxena and Deepali Rana and Ayesha Varshney and Sahinur Rahman Laskar}, year={2026}, eprint={7577341}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2605.10155 }