VERDICT AI

Walk into court already prepared. A multi-agent AI courtroom that simulates opposing counsel, flags objections instantly, and generates attorney-ready reports from voice and facial signals.

Comment

TRY OUT LINK: sarah.chen@demo.com Demo!Pass123

Inspiration

Legal witness preparation is still highly manual, expensive, and inconsistent. Mid-sized law firms spend significant billable hours preparing clients for depositions and trials, yet witnesses often enter the courtroom underprepared and stressed.

We asked a simple question: What if every witness could face opposing counsel before the real courtroom?

VERDICT AI was inspired by the gap between how high-stakes courtroom performance is evaluated and how little technology exists to simulate it. We set out to build an AI-native, multimodal courtroom environment that transforms witness prep into a measurable, repeatable, and data-driven process.

What it does

VERDICT AI is a multi-agent AI courtroom simulator that stress-tests witnesses before trial. It recreates adversarial questioning, flags objections in real time, and generates attorney-ready reports that lawyers can use to coach clients effectively.

Core capabilities: • Simulates realistic opposing counsel questioning • Flags evidentiary objections in real time using RAG grounded in the Federal Rules of Evidence • Performs facial recognition and audio-based witness evaluation • Analyzes witness behavior using voice, facial, and linguistic signals • Generates structured, attorney-ready performance reports after each session

The result is a practice environment that measures not just what a witness says, but how they perform under pressure.

How we built it

VERDICT AI is powered by a coordinated multi-agent architecture and a retrieval-grounded legal intelligence layer.

Multi-Agent System • Interrogator Agent: Generates adversarial, context-aware questions to mimic real opposing counsel • Objection Copilot: Uses a RAG pipeline and Finetuned Nemotron Model over Federal Rules of Evidence to flag objectionable questions in real time • Behavioral Analysis Agent: Performs multimodal analysis using facial recognition and audio evaluation • Review Orchestrator: Produces structured, attorney-ready performance reports

AI and Infrastructure Stack • Nemotron — fine-tuned on legal questioning patterns to improve adversarial tone and deposition realism • Claude SDK — used for agent orchestration, reasoning workflows, and system coordination • ElevenLabs — provides realistic voice synthesis for opposing counsel simulation • Lovable — enabled rapid full-stack development and frontend experience • Nia — powers the semantic retrieval layer for legally grounded responses • Databricks — handles secure data processing, storage, and pipeline orchestration

We also built a custom eval framework to systematically measure objection accuracy, questioning realism, and behavioral scoring quality during development.

Challenges we ran into

Real-time latency

Objection detection and agent responses must feel instantaneous in a live deposition setting. We optimized streaming pipelines and retrieval paths to maintain low-latency performance.

Multimodal synchronization

Aligning facial signals, audio features, and transcript timing required careful normalization and buffering to avoid noisy or misleading behavioral scores.

Legal grounding vs. speed

Large legal corpora are heavy to index in a hackathon environment. We curated a high-signal subset of the Federal Rules of Evidence and procedural materials to balance accuracy with fast ingestion.

Adversarial realism

Generic LLMs tend to be cooperative. We fine-tuned Nemotron and separated agent roles to produce more authentic opposing counsel behavior.

System evaluation

Legal AI requires trust. We built custom evals to continuously validate objection detection accuracy and questioning quality.

Accomplishments that we're proud of

• Built a fully working multi-agent AI courtroom simulation in hackathon time
• Achieved real-time objection flagging grounded in legal authority
• Integrated multimodal intelligence (facial recognition + audio evaluation + speech analysis)
• Fine-tuned Nemotron for adversarial legal questioning
• Developed a custom eval framework for legal AI reliability
• Delivered attorney-ready structured reports, not just chat outputs
• Orchestrated a production-style stack using Lovable, Claude SDK, Nia, ElevenLabs, and Databricks

Most importantly, we transformed witness preparation from subjective coaching into measurable performance intelligence.

What we learned

• Multi-agent architectures dramatically improve realism in interactive legal AI
• Multimodal signals reveal coaching insights lawyers actually care about
• In legal workflows, structured reports drive adoption more than chat interfaces
• Fine-tuning is critical for adversarial tone and courtroom authenticity
• Systematic evals are essential for trustworthy legal AI systems
• Latency and grounding matter more than raw model size in live simulations

We also learned that the future of legal preparation is simulation-driven, continuous, and data-backed.

What's next for VERDICT AI

• Jurisdiction-specific rule packs and expanded legal coverage
• Longitudinal witness improvement tracking across sessions
• Deeper behavioral risk and inconsistency detection
• Firm-specific fine-tuning and secure case isolation
• Expanded deposition and trial scenario libraries
• Enterprise-grade analytics dashboards for litigation teams

Our vision is to make AI-powered courtroom simulation the standard preparation layer for modern litigation teams.

Built With

  • alembic
  • alembic-frontend:-react-(vite)
  • amazon-web-services
  • aws-s3-authentication-&-security:-jwt-(httponly-cookies)
  • azure
  • claude-sdk
  • databricks
  • delta-lake
  • elevenlabs
  • face-recognition
  • facial
  • facial-recognition-(mediapipe-/-facs)-data-&-infrastructure:-nia-(semantic-retrieval)
  • facs
  • fastapi
  • github
  • javascript
  • javascript-backend-&-apis:-fastapi
  • jwt-(httponly-cookies)
  • lovable
  • lovable-voice-&-media:-elevenlabs
  • mediapipe
  • multimodal-analysis-(voice
  • nemotron-(fine-tuned)
  • nia-(semantic-retrieval)
  • okta
  • postgresql
  • python
  • react-(vite)
  • real-time-audio-processing
  • real-time-streaming
  • redis
  • rest-apis
  • retrieval-augmented-generation-(rag)
  • saml
  • saml-2.0-(okta-/-azure-ad)-developer-tools:-github
  • speech)-languages:-python
  • speech-recognition
  • sqlalchemy
  • typescript
  • voice-recognition
  • vscode
  • websockets
  • websockets-(real-time-streaming)
Share this project:

Updates