Towards “Zen” Generative AI
Your LLM under control.
Two lines of code.
Zero added latency.
UncleJean judges every answer your AI gives in production — hallucinations scored, dangerous answers blocked, every token accounted for.
GDPR-native · Built in France · Sovereign by design
client = UncleJean(OpenAI())
One line changed. Nothing else. Analysis runs off the request path: your users never wait for us — and our outage can never become yours.
The problem
Software is deterministic. Your LLM is not.
The old world monitored 404s and latency. The new world must watch hallucinations, toxicity, sycophancy — and silent model updates you were never told about. Most enterprises run an LLM in production with the old dashboard. That gap is where reputations die.
How it works
From install to insight in minutes
Connect
pip install unclejean + your API key.
Five minutes to your first trace — the bar we set ourselves, timed live in every demo.
Score
Every answer evaluated against the question that produced it — the Trustworthiness Index, your AI’s weather forecast. Hallucination, toxicity and drift, measured continuously.
Act
The Circuit Breaker throttles, blocks or substitutes a dangerous answer — before the damage is done, with human-in-the-loop supervision.
The platform
Four pillars, one trust layer
Output Monitoring
Hallucination, toxicity and sycophancy detection — a Trustworthiness Index for your AI.
Operational Safety
Real-time circuit breaker on the answer. Throttle, block or reroute — before damage is done.
FinOps
Token spend per model, per team, per prompt version — and the waste it surfaces: retries, runaway agent loops, bloated context.
Versioning
Full traceability of prompts, RAG pipelines and agents.
Built in France. Sovereign by design.
Design partners
The pilot costs you nothing. Ever, if we’re wrong.
We take a handful of design partners. The pilot is free — and if we surface zero issues and zero savings in your LLM traffic within 30 days, you owe us nothing. Ever.
Two lines of code to start. Your data stays yours — never used for model training.
Who’s behind this
No badge wall. Just receipts.
We build in the open on Hugging Face — evals resources and the Golden Dataset method.
Organizers of AI Day Chevreuse — next edition October 24, 2026.
30 years running production-grade delivery for large companies in rail transport, retail and banking.
Defense·Aerospace·Legal·Finance & Insurance·Healthcare