DealProof

Inspiration

Client agreements rarely live in one place. Pricing may be agreed in email, deadlines changed in chat, and payment terms confirmed somewhere else. When AI is then asked to draft the contract, it can produce something that sounds perfectly reasonable while silently changing a critical term.

We built DealProof to make AI-generated contracts traceable to what was actually agreed.

AI can draft it. DealProof proves it.

What it does

DealProof turns client conversations into an evidence-backed Agreement Ledger.

It extracts commercial terms, resolves negotiation history, and connects critical terms to their source evidence. Human reviewers can explicitly approve terms that require review.

From the reviewed ledger, DealProof generates a structured contract while preserving provenance between contract values and their underlying terms.

Before finalization, a separate verification pass checks the generated draft against the resolved agreement. If a critical value is contradictory, unsupported, or missing, a deterministic policy gate blocks the contract from being finalized.

In our demo, the parties agree on a $12,000 project fee. We simulate an AI drafting error that changes it to $15,000. DealProof detects the contradiction and blocks finalization.

How we built it

DealProof uses a Next.js/React frontend with a FastAPI backend.

Google Gemini powers AI-assisted extraction, negotiation analysis, contract generation, and verification. Structured agreement terms retain provenance and evidence references throughout the workflow.

The pipeline is:

Source conversation → Agreement Ledger → Human review → Contract generation → Verification → Deterministic policy gate

The final decision to allow or block finalization is enforced by deterministic application logic rather than being delegated to the language model.

Challenges we ran into

One of the hardest problems was distinguishing a term that was merely mentioned from one that was actually accepted. Negotiations contain proposals, counter-proposals, rejected values, and superseded terms.

Another challenge was preserving provenance through contract generation so that generated values could still be traced back to their agreement terms and evidence.

We also had to design the system so AI could assist with reasoning without giving the model authority over whether a contract was safe to finalize.

Accomplishments that we're proud of

• Evidence-backed Agreement Ledger
• Negotiation-state resolution
• Human approval workflow for uncertain terms
• Provenance-preserving contract generation
• Verification of generated contracts against agreed terms
• Detection of contradictory and unsupported critical terms
• Deterministic blocking of unsafe contract finalization
• End-to-end working demo

What we learned

The biggest lesson was that trustworthy AI systems need more than better prompting. Probabilistic models are powerful for extracting, interpreting, and generating information, but critical business decisions benefit from explicit evidence and deterministic safeguards.

We treated provenance as part of the product itself rather than hidden metadata.

What's next for DealProof

Next, DealProof could connect directly to email, Slack, CRMs, and document workflows so agreement evidence can be captured automatically. We would also expand the commercial-term schema, strengthen negotiation resolution, and add production-ready collaboration and contract export integrations.

Built With

• docker
• fastapi
• gemini
• llm
• nextjs
• postgresql
• pydantic
• python
• react
• render
• restapi
• typescript
• uvicorn