Formal Methods × Generative AI
From AI you hope is right
to AI you can prove is right
A plug-and-play layer that verifies every reasoning step of your LLM against your rules, in real time. Any model, no retraining.
Get in touchWhat you get
Every step verified against your rules
Example: an AI assistant helping engineers choose materials for an aircraft part, verified against the company’s engineering rules
Without Formel AI
With Formel AI
How it works
A formal layer around your LLM
1 · Plug in
Any LLM, your infrastructure
Works with any open or closed-source LLM, no retraining. Our layer itself is CPU-only: one container, deployable in any cloud (AWS, GCP, Azure…), on-premise or air-gapped.
2 · Verify
Logic rules + your rules
Logic rules catch contradictions from day one. Your rules (regulation, internal policy, engineering constraints) are written in plain language and set up in days.
3 · Prove
An answer you can audit
Every answer comes with a reasoning certificate: a graph of each assumption and deduction. A step that can’t be verified is flagged, never silently passed.
Business impact
From “we can’t put AI in production”
to “we can prove it”
Fewer tokens and fewer people: smaller models, no judges, no human review
Use cases that were too expensive become profitable
Every step is verified or flagged, never passed unchecked
Incidents, rework and escalations avoided
Every answer ships with the proof of its reasoning
Unlocks defense, critical infrastructure and regulated sectors
In practice
| Without Formel AI | With Formel AI | |
|---|---|---|
| Model size | LargeModels are over-sized to compensate for unreliability | SmallerReliability comes from the formal layer |
| LLM-as-a-judge | 10+One judge is needed per business rule | 0Rules are verified formally instead |
| Manual review | 10+ peopleThey review 10,000+ answers/month | NoneFormel AI flags the wrong ones |
Who benefits
Built for every stakeholder
Tech Lead / Product Manager
Engineering & Product
- Domain experts write each rule once, in plain language
- Rules are easy to update if requirements change
- No retraining, no change to your existing agents
CISO / Compliance Officer
Security & Compliance
- Runs in your cloud or on-premise, even air-gapped
- A reasoning certificate for every answer, ready for auditors
End user
Trust & Adoption
- Trust the answer without re-checking it
- See why the AI reached its answer
Who we are
Meet the team
Co-founders
Prev. Senior ML Product Engineer at Probabl, the Inria spin-off and official operator of scikit-learn. Former data science lecturer at Ecole polytechnique (EXED and X-HEC Data Science for Business MScT) and CentraleSupélec EXED. PhD from ENS Paris-Saclay.
Prev. ML Engineer at Qubit Pharmaceuticals, after working on the company’s HPC calculation management as a Software Engineer. PhD from ENS Paris-Saclay, MSc from Ecole polytechnique and MVA (Mathématiques, Vision, Apprentissage).
Founding team
Prev. Social Commerce Lead at Primelis, VC Analyst at Eurazeo and Plug & Play. MiM from ESCP.
Prev. Data Science Lead at Qubit Pharmaceuticals. PhD from Université Paris Cité, MSc from ESPCI and Sorbonne Université.
Prev. postdoc at Inria and ENS Paris-Saclay (Deducteam). PhD and MSc from Heriot-Watt University.
Prev. Head of ML at Gorgias, Lead ML Engineer at Agicap. PhD from ENS Paris-Saclay, MSc from Ecole polytechnique and ENSTA.
Prev. Senior Software Engineer at Nomadic Labs, Alstom, and Tweag. PhD from ENS Paris-Saclay (Deducteam) and Mines Paris, MSc from Ecole polytechnique.
Prev. AI Engineer at BNP Paribas and Full-Stack Engineer at Qubit Pharmaceuticals. MSc from Ecole polytechnique and ULB.
Prev. ML Engineer at Gorgias. MSc from Oxford, BSc from EPFL.
Business angels
and many more…