Use cases
Functional AI demonstrators, live and testable — from framing to production.
- Documents & RAGAssurConseil RAGAgentic RAG assistant for insurance contract interpretation, with a Claude / Mistral comparison in document mode.
- Compliance & RiskFraud AgentPre-assessment agent for P&C insurance claims: business rules set the routing, a Claude agent proposes, a claims handler decides.
- Governance & FramingAGAPAI-arbitrated project portfolio prioritization, free of Business/IT bias.
- Documents & RAGLexGuardDetects red-flag signals in contracts before signature.
- Governance & FramingComexStrategic framing and alignment through AI.
- Public DataSireneNatural language company search.
- Public DataDVFReal estate transaction search and analysis.
- Compliance & RiskID Document VerifierAI-powered document quality control and MRZ consistency check.
⏳ Slower first load: to keep costs down, the demo goes to sleep when no one is using it. On first access, the server restarts; after that, it responds normally.
AssurConseil RAG
Agentic RAG assistant for insurance contract interpretation, with a Claude / Mistral comparison in document mode.
What It DemonstratesIndustrialising a RAG: evaluation before production, cost guardrails, proven European inference, progressive deployment.
Business Challenge
Insurance advisors must navigate dozens of pages of general terms to answer client questions accurately, under regulatory duty-to-advise obligations — a misread clause or wrongly cited guarantee exposes the insurer to liability.
AI Solution
An agentic RAG assistant that routes each question to the most suitable pattern (ReAct for direct lookups, Plan-and-Execute for multi-part questions), with mandatory source citation and explicit refusal to answer when no supporting document exists — a demo built on fictional contracts and a fictional company (ARESIA Assurances), illustrating an agentic architecture for regulated advisory use cases. In document mode, the same question is sent in parallel to Claude and Mistral Large 3 (Mistral inference run in the EU): answers side by side, with latency, tokens and cost.
Key Features
Adaptive ReAct / Plan-and-Execute Routing
A router detects question complexity and switches between direct retrieval (ReAct) and sub-question decomposition (Plan-and-Execute).
Anti-Hallucination Guardrail
Mandatory citation of source articles; an explicit "not found" response rather than a fabricated answer when information is absent from the documents.
Full Reasoning Traceability
Every graph step (router, retrieve, evaluate, synthesize) is logged in an inspectable trace_log, from question to final answer.
Your Contract, Two Engines
Upload a contract: temporary in-memory index, separate from the corpus. The same question goes in parallel to Claude and Mistral Large 3 (EU inference); answers side by side with latency, tokens and estimated cost.
Tech Stack
Fraud Agent
Pre-assessment agent for P&C insurance claims: business rules set the routing, a Claude agent proposes, a claims handler decides.
What It DemonstratesDesigning a bounded agent for a regulated business: what the code decides, what the agent proposes, what the human approves, with measured cost and latency.
Business Challenge
In a claims department, most cases are simple, but those that call for an investigation or a recovery claim against a liable third party require cross-checking several documents. Handing this work to AI raises three questions: who decides the routing, what happens when the model gets it wrong, and what does it cost?
AI Solution
On fictional claims, business rules compute a score out of 50 and set the routing (straight-through processing, investigation, fraud-unit alert). The agent cannot change it. A Claude agent is called only when useful, when the case is suspicious or a police report is provided. It picks its tools, returns a structured proposal, then stops: a claims handler approves or rejects it with a reason, before the agent drafts a letter that is never actually sent. Thresholds are demonstration choices, not calibrated on real data.
Key Features
Rules first, AI when useful
Score and routing come from deterministic rules. A clean case without a police report triggers no AI call.
An agent that picks its tools
Depending on the case, the agent checks the repair shop's history or computes the recovery amount. Every step shows in a log (tool, input, output, duration), and amounts are recomputed by the code.
Mandatory human approval
The handler approves, or rejects with a reason. A recovery letter draft and a D+5 / D+15 / D+30 plan follow, marked "never actually sent".
Measured figures
- $0.02-0.03 per analysis
- 10-20 s
- 15 live analyses per day, then a labelled replay (measured on 01/10/2026)
Tech Stack
AGAP
AI-arbitrated project portfolio prioritization, free of Business/IT bias.
What It DemonstratesTooling a portfolio decision: surfacing Business/IT gaps and arbitrating through scenarios, with tracked justifications.
Business Challenge
A Portfolio Director arbitrates dozens of projects every year based on structurally biased Business and IT declarations: Business overstates business impact, IT overstates effort and technical risk to protect itself. Without a mediation mechanism, final arbitration becomes a political power struggle rather than an objective decision.
AI Solution
A governance application that has a Claude agent separately challenge each Business and IT declaration, computes a confidence index and an adjusted attractiveness score, then visualizes perception conflicts before final arbitration — a demo built on a fictional 6-project portfolio, illustrating neutral AI mediation between two parties with diverging interests.
Key Features
Multi-Role AI Challenger
A Claude agent separately interrogates Business (perceived value) and IT (estimated effort), generates clarifying questions, and computes a confidence index per declaration.
Conflict Map
Visual mapping of the Business/IT perception delta, with a conflict threshold configurable at the deployment level (currently 20 points) to flag projects needing mediation before arbitration.
Scenarios & Arbitration Cut-Line
Side-by-side project ranking under two scoring scenarios (MVP / Full) with an adjustable cut-line, plus forced inclusion always justified and logged.
Tech Stack
LexGuard
Detects red-flag signals in contracts before signature.
What It DemonstratesMaking a contract readable for a non-lawyer: red-flag signal detection and a plain-language verdict.
Red-flag signal detection
Identifies patterns of risky contractual arrangements.
Plain-language verdict
3-level assessment, no legal jargon.
No application-side storage
The document is processed in memory and is not stored by the application. The AI provider (Anthropic) retains transmitted data for 30 days, under its default policy.
Comex
Strategic framing and alignment through AI.
What It DemonstratesStructuring AI framing: from questionnaire to two AI-costed implementation paths.
Framing + AI Clarification
19 structured questions, followed by an AI clarification exchange (up to 4 targeted questions) to refine the diagnosis.
Two Costed Proposals
A "quick" proposal (under 3 months, existing stack) and a "balanced" proposal (3-6 month MVP), with budget, timeline and risks estimated by the AI.
Tailored to Your Environment
Proposals take the declared IT stack into account (cloud and AI provider already in place).
Sirene
Natural language company search.
What It DemonstratesMaking a public API accessible in natural language, with a lightweight model (Claude Haiku).
Natural Language Queries
Translation of the query into Sirene API parameters by an LLM (Claude Haiku).
Sector and Region Filters
NAF sector and region inferred from free-form expressions, with a filter on administrative status (active or discontinued companies).
Results Export
CSV and Excel export of results.
DVF
Real estate transaction search and analysis.
What It DemonstratesUsing open land-registry data: address search and municipality-level price statistics.
Median Land Price by Municipality
Median price per m² and land area, with year-by-year evolution, calculated for each municipality from DVF transaction data.
Address-Based Transaction Search
Geolocated search for real estate and land sale transactions around an address, with automatic fallback to a wider radius.
ID Document Verifier
AI-powered document quality control and MRZ consistency check.
What It DemonstratesCombining AI reading with a deterministic calculation (MRZ check digits) to make verification more reliable.
Instant quality control
Detection in seconds of blurry photos, glare, or partially out-of-frame documents, before a file ever reaches a handler.
Verifiable MRZ consistency
Deterministic ICAO 9303 check-digit calculation (ID card and passport), with the details shown. The characters are read by the AI: the calculation is transparent, the reading can be wrong.
Front/back cross-check
Comparison of the document number between both sides of the card, as an additional consistency signal.