AI R&D strike team for hard decisions under uncertainty
Jolibrain helps engineering teams tackle consequential problems whose feasibility is still uncertain - where a standard product, routine fine-tune, or commodity integrator is not enough.
We formulate the decision, test the riskiest assumptions first, stop weak approaches early, and build the ones that survive - from prototype to integration and transfer.
Bring us a hard decision →who we are
We are a small team of PhDs and senior engineers in Toulouse. We work directly with R&D and engineering organizations on important problems whose technical route is not yet obvious.
Like a Miles Davis band, we value competence, originality, collaboration, and room to improvise. Senior people do the work. We stay method-agnostic, candid about uncertainty, and focused on evidence rather than AI hype.
where we help
Critical decisions from evidence. Detect, reject, alert, inspect, localize, validate, or proceed from imperfect visual, sensor, or documentary evidence.
Evidence for rare events. Create the missing observations or detail required to train and test a system, while preserving the physical or semantic meaning that matters downstream.
Multi-step decisions. Select robust action sequences while durations, resources, visibility, failures, interactions, and future states evolve.
We combine machine learning, reinforcement learning, optimization, simulation, classical methods, and software engineering as the problem requires.
selected evidence
Critical decision systems - DeepDetect. Jolibrain's open-source platform for self-hosted AI systems, from rare-event catenary monitoring and aircraft inspection to satellite-telemetry anomalies, small-target detection, and vision-based landing for civil and military aircraft.
Missing evidence - JoliGEN. Jolibrain's open-source generative toolkit for filling data gaps with rare and long-tail cases, label-preserving sim-to-real evidence, and decision-useful super-resolution validated on downstream engineering tasks.
Technical knowledge - Colette. Open-source, self-hosted multimodal retrieval that preserves the tables, figures, diagrams, and page structure engineers actually use.
Planning under uncertainty - Wheatley. Public research and open-source software for learned scheduling under resource constraints and uncertain durations.
how we work
Start with the decision. We review what the system must decide, what is uncertain, which errors matter, what evidence exists, which baselines are credible, and what would justify continuing or stopping.
If feasibility survives, a small senior team attacks the highest-risk technical hypothesis, then builds toward a prototype, integration, early production, and transfer.
A useful result may be a working system or clear negative evidence that prevents a much larger dead end.
Bring us the decision your current approach cannot make reliably →open source
We maintain open-source foundations including DeepDetect, JoliGEN, Colette, and Wheatley alongside confidential client systems.
"Machines aren't the thing,
Joe McMillan, "Halt and Catch Fire"
they're the thing that gets us to the thing"
We're based in Toulouse, France, and we like it like that.
