LegalOS

home page
pdf upload
example of dispute simulation

Inspiration

77% of Americans face at least one legal issue per year, but most can't afford the $300–500/hour it costs to have a lawyer read their own contracts. People sign bad deals, settle winnable cases, and walk into negotiations blind — not because they're wrong, but because they can't afford to be right. We built LegalOS to change that.

What it does

Upload any legal document — a contract, complaint, or court filing — and LegalOS gives you:

Clause-by-clause risk analysis — every clause scored HIGH/MEDIUM/LOW with a 0–100 risk score and plain-English explanation
Overall danger score — one number that tells you how risky the document is
Precedent case lookup — real US court cases relevant to each clause
Dispute simulation — Monte Carlo-style litigation prediction with win probability, cost estimates, and a FIGHT/SETTLE/NEGOTIATE recommendation
Live AI negotiation call — practice arguing against AI opposing counsel who defends the contract in real time with voice, responds to your specific objections, and coaches you on what to say
Pitch trainer — record your objections and get scored on confidence, clarity, and legal strength

How we built it

Frontend: Next.js 14 (App Router) with Tailwind CSS, Framer Motion for animations, and Recharts for simulation visualizations.

AI/LLM: OpenAI GPT-4o-mini powers all legal reasoning — clause analysis, risk scoring, dispute simulation, opposing counsel responses, precedent search, and plain-English explanations. All through Next.js API routes.

Voice: ElevenLabs text-to-speech (eleven_multilingual_v2 model) gives opposing counsel a realistic voice during live negotiation calls. Web Speech API handles real-time speech-to-text for user input. Falls back to browser speech synthesis if ElevenLabs is unavailable.

Data layer: MongoDB Atlas stores document history and analysis results. We chose MongoDB's document model because legal data is naturally nested — a document contains clauses, each clause has scores, explanations, and precedent cases.

Enterprise intelligence: Palantir Foundry integration via the AIP Actions API. Every analysis syncs 6 ontology object types to Foundry — Legal Documents, Clauses, Risk Analyses, Dispute Simulations, Parties, and Court Cases — enabling cross-document queries and dashboards at scale.

Infrastructure: Deployed on AMD EPYC-powered cloud infrastructure through Vercel (serverless functions) and Render (original Python ML backend). AMD handles the burst compute pattern of our document processing pipeline.

Auth: Clerk for authentication and route protection.

Challenges we faced

ElevenLabs model deprecation — eleven_monolingual_v1 stopped working mid-hackathon and we had to migrate to eleven_multilingual_v2 and add browser speech fallback
Getting GPT to return consistent structured JSON for legal analysis across wildly different document types (contracts vs complaints vs court filings)
Building the live call state machine — coordinating speech synthesis, speech recognition, API calls, and UI transitions across 6 call phases without race conditions
Palantir ontology design — figuring out the right object types and relationships to make cross-document queries meaningful

What we learned

Legal documents are surprisingly varied — a contract risk engine breaks immediately on court filings unless you rethink what "risk" means for each document type
Voice UX is hard — users need to see what to say while they're speaking, not just before
Fire-and-forget API patterns (like our Palantir sync) are essential when you don't want third-party latency blocking the user