Sentinelle
Contract decoder & negotiation copilot — the AI that defends the signer: sorted red flags, a risk verdict and sourced answers in under two minutes
Discuss a similar projectThe live site
Click the preview to open and test Sentinelle directly in your browser
sentinelle-app-production.up.railway.app
Réponse courte
Sentinelle is an AI-powered contract decoder built by Digit-AI during the Vibe Coding Arena 2026 hackathon (the 'FinTech' brief). Aimed at any signer without access to a lawyer (individuals, micro-businesses, freelancers), it doesn't just summarize: it defends the signer. From a PDF or pasted text, it splits the contract into clauses, detects and sorts red flags by severity, delivers an overall risk verdict and answers questions by citing the exact text — all in under two minutes. It quantifies the impact of clauses (€/year), drafts a negotiation email and cites the real article of law. The analysis relies on Claude Opus 4.8 with mandatory verbatim citation (anti-hallucination) and an analysis cache guaranteeing the reliability of the demo.
We sign what we don't read
Insurance, loan, terms of sale, lease: without a lawyer, the signer faces an information asymmetry and misses the trap clause
Individuals, micro-businesses and freelancers sign multi-page contracts without access to a lawyer. They miss the auto-renewal clause, the penalty, the cap that's too low — and find out too late. Sentinelle removes this information asymmetry: the tool defends the signer, quantifies the impact of risky clauses and cites the real article of law, citation in hand.
Information asymmetry
The contract's author knows its traps; the signer doesn't, and has no lawyer to rebalance things.
Missed trap clauses
Auto-renewal, penalties, low caps: costly clauses buried in the legalese.
No negotiation leverage
Even when spotting a clause, the signer doesn't know how to quantify it or which article of law to cite.
Risk of AI hallucination
A generic assistant might invent a clause or a law — unacceptable on a legal topic.
Functional scope
Individuals, micro-businesses and freelancers facing a contract (insurance, loan, terms of sale, lease). Processes covered: PDF/text ingestion, splitting into clauses, flag detection and sorting, risk verdict, sourced Q&A, impact simulation and a negotiation copilot.
Key constraints
Built in 24h (hackathon) · 100% Railway deployment (web + FastAPI + Postgres) · Systematic citation of the text (anti-hallucination) · Ephemeral processing, no persistence · 'Reading aid, not legal advice' disclaimer.
What Sentinelle solves
Three axes to understand a contract, secure it and negotiate it — without a lawyer
Understand in under 2 minutes
Turn 14 pages of legalese into red flags sorted and explained in plain language, with an immediate risk verdict.
Decide with confidence
Every flag and every answer cites the text verbatim; no unsourced statement, to eliminate the risk of hallucination.
Move to negotiation
Quantify the impact (€/year), generate a ready-to-send email and cite the real article of law to rebalance the power dynamic.
What Sentinelle enables
From contract ingestion to negotiation, a complete chain in the signer's service
PDF or text ingestion
Upload a PDF or paste the text directly, with extraction then conditional OCR if the text layer is insufficient, and splitting into clauses.
Red flag detection
Spotting risky clauses (auto-renewal, penalties, low caps…), sorted by severity and explained in plain language for the general public.
Overall risk verdict
Aggregation of the flags into a clear verdict (maximum severity), to tell at a glance whether the contract deserves particular vigilance.
Sourced Q&A on the contract
Natural-language questions with answers citing the exact contract text (clause_id); the front end highlights the relevant clause.
Impact simulation & negotiation
Quantified estimate (€/year) of risky clauses and generation of a ready-to-send negotiation email, with fallback templates.
Légifrance citation & deadline
Citing the real article of law (a set of pre-loaded key articles) and generating an .ics file so you never miss a termination date.
Tech stack
A React / FastAPI full-stack deployed on Railway, built around a central JSON schema and citation traceability
Backend
Frontend
Artificial intelligence
AI robustness
Data & cache
Infrastructure
Differentiating pattern — Verbatim citation & central JSON schema: the analysis runs in a single structured LLM pass (clauses → flags → severity → citations) validated by schema, with a mandatory verbatim citation for every statement — rejected if absent from the source text. Since the contract fits in the 200k-token window, no vector database (RAG) is needed, and prompt caching reuses the contract at ~0.1× the cost on each question.
Architecture diagrams
A Railway full-stack: a React SPA that highlights clause by clause, a FastAPI API that orchestrates the Claude analysis and a Postgres cache for demo reliability
User
Signer (individual, micro-business, freelancer)
Uploads a PDF or pastes the text
Web interface
React + Vite SPA
Rendering highlighted by clause_id · streamed Q&A area
Analysis API (FastAPI)
FastAPI + uvicorn (uv)
Ingestion, analysis, Q&A, verdict
Central JSON schema
Clauses · flags · severity · citations
AI core
Claude Opus 4.8 (analysis)
'Advocate for the signer' stance
Claude Sonnet 4.6 (live fallback)
Latency / cost
Prompt caching
Q&A at ~0.1× the cost
Persistence & cache
Railway Postgres
Hash-based analysis cache + demo fixtures
What the prototype demonstrates
Indicators from the prototype presented at the Vibe Coding Arena 2026 hackathon
| Indicator | Without Sentinelle | With Sentinelle | Gain |
|---|---|---|---|
| Understanding a contract | 14 pages of legalese left unread | Red flags explained in plain language | Accessible |
| Detecting trap clauses | Auto-renewal, penalties missed | Detection + sorting by severity | Secured |
| Tool stance | Neutral summary | Defends the signer, quantifies the impact | Pro-signer bias |
| Answer reliability | Risk of hallucination | Systematic verbatim citation | Verifiable |
| Negotiation capability | No leverage without a lawyer | Email + enforceable article of law | Equipped |
Qualitative gains
The signer's advocate
The tool doesn't summarize: it takes the side of the person signing, quantifies the impact of clauses and cites the real article of law.
Anti-hallucination
Every statement cites the text verbatim; a citation absent from the document is rejected. Trust rests on the source, not the tone.
Confidentiality
Ephemeral processing of contracts, with no production persistence: a trust advantage for sensitive documents.
Controlled latency
Content-hash analysis cache and contract prompt caching: repeated questions cost ~0.1× and respond almost instantly.
Graceful degradation
Each extension (simulation, negotiation, Légifrance, .ics) has a deterministic fallback: its failure never breaks the analysis core.
Responsible framing
An explicit 'reading aid, not legal advice' disclaimer and systematic citation, for clear and honest use.
How about we start by talking?
A similar project in mind? Let's take 30 minutes to understand your situation and see how AI can help. First conversation free and with no commitment.
Discuss a similar project