NyayaSetu – Devpost Submission Draft

Multi-Agent Legal Co-Counsel | Nexora Hacks 2026

1. Problem Statement & Motivation

Justice in India is structurally inaccessible for a large section of citizens.

Legal consultation costs ₹2,000–₹15,000 per hour.
Per capita income is ~₹14,000 per month.
Legal language (“Legalese”) is complex and intimidating.
Over 5 crore cases are pending in courts.
Laws may be passed but temporarily not enforced (e.g., provisions kept in abeyance), leading to confusion.

For the common citizen, justice becomes a luxury instead of a right.

NyayaSetu was built as a zero-cost, instant “Legal First Responder” — designed to bridge the gap between citizens and the legal system using AI.

2. Our Solution: What NyayaSetu Does

NyayaSetu is a Multi-Agent Legal Co-Counsel system that doesn’t just retrieve laws — it reasons over them in real-world context.

Core Capabilities

1. See – Multimodal Evidence Analysis

Accepts video/audio evidence (e.g., CCTV footage).
Uses multimodal AI to extract factual observations.
Generates structured legal-relevant summaries from raw media.

2. Read – Legal Retrieval

Maps layperson queries to relevant IPC/BNS sections.
Uses hybrid semantic + keyword search for precision.
Explains statutes in simplified language.

3. Verify – Live Enforcement Check

Cross-references live web sources and court updates.
Detects if a law is in force, amended, or kept in abeyance.
Prevents “technically correct but legally outdated” advice.

3. How It Works (Architecture Overview)

We moved beyond basic RAG and implemented an Agentic Workflow using LangGraph.

Agentic System Design

1. Legal Clerk

Retrieves relevant statutes.
Uses Qdrant Hybrid Search.
Converts user queries into structured legal mappings.

2. Evidence Auditor

Analyzes uploaded video/audio.
Uses Gemini 2.5 Flash for multimodal reasoning.
Extracts legally relevant observations.

3. Amendment Watchdog

Performs live browsing via DuckDuckGo.
Cross-checks Supreme Court judgments and news.
Detects enforcement status and legal updates.

4. Senior Counsel

Synthesizes:
- Evidence findings
- Statutory provisions
- Live enforcement updates
Produces grounded, citation-backed advice.

4. Design Philosophy: “Paper & Ink” Trust Layer

We intentionally rejected the futuristic “cyberpunk AI” aesthetic.

Legal counsel requires trust, familiarity, and psychological comfort.

Design Choices

Color System
- Official Maroon (#800000) → Authority
- Parchment Beige (#FFFBF0) → Reduced eye strain
Typography
- High-contrast serif fonts (e.g., Merriweather)
- Improves readability of dense legal text
- Optimized for elderly users (“Grandmother Test”)
Skeuomorphic Trust
- Interface mimics a physical legal pad.
- Reduces tech intimidation.
- Encourages adoption among non-digital-native users.

5. Technologies & Tools Used

LangGraph – Agent orchestration
Qdrant – Hybrid vector search for legal retrieval
Gemini 2.5 Flash – Multimodal forensic analysis
DuckDuckGo Search – Live web cross-verification
Python – Backend orchestration
Frontend Stack – Custom UI with trust-focused design
Vector embeddings & hybrid semantic retrieval pipeline

6. Key Challenges & How We Solved Them

1. Hallucination Risk

Problem: AI inventing court cases or judgments.
Solution:

Mandatory citation enforcement.
URL-backed answers only.
Automatic disclaimer fallback if verification fails.

2. Search API Fragility

Problem: Live web search timeouts during demos.
Solution:

“Demo-Safe Fallback” layer.
Cached registry for critical legal topics.
Graceful degradation instead of system failure.

3. Abeyance Detection

Problem: Law may be passed but not enforced.
Solution:

Built logic to distinguish:
- Enacted
- Amended
- Stayed
- In abeyance
Integrated live court update validation.

7. Accomplishments

Successfully detected the abeyance status of controversial provisions (e.g., hit-and-run law).
Built a resilient multi-agent system that survives API failures.
Implemented native video evidence analysis without external OCR dependencies.
Created a legally trustworthy AI flow with citation enforcement.

8. Functional Project Status

Working prototype with:
- Legal query interface
- Video upload analysis
- Live verification layer
- Structured legal advice output
Built during Nexora Hacks 2026 hackathon period
Software-only project (extensible to hardware evidence capture)

9. Demonstration Video

Youtube Video - Attached Below

10. Source Code & Repository

GitHub Repository: Attached below
Includes:
- Setup instructions
- Environment configuration
- Run commands
- Architecture explanation
Repository can be public or shared privately with organizers upon request.