We work on AI systems that have to run in real conditions. Most of the time the model is fine and the data is not. We start by finding out which.
We map the pipeline end to end: sources, labels, splits, metrics, deployment constraints. You get a written account of where performance breaks and why.
Rebuilding a training set so it matches the distribution you actually deploy into. Gap analysis, relabelling, balance measurement.
Diffusion and simulation for cases you cannot photograph: rare defects, accidents, edge geometry. Published methods, not guesswork.
VLM and LLM annotation with human review where it matters. Built to cut annotation cost by an order of magnitude, then measured.
Training, fine-tuning, distillation and honest evaluation. We report the failure modes as well as the headline number.
Getting a model onto constrained hardware and keeping it there: quantisation, pruning, stereo rigs, field monitoring.
Two weeks inside your system. No commitment beyond it.
A fixed plan with the metric we will move and how it gets measured.
Weekly checkpoints, code in your repository from day one.
Your team runs it without us. We stay reachable.
Some of this was done inside Trusted AI Labs at UMONS, where the founding research took place.
An NVIDIA Jetson stereo vision system watching active rail worksites for hazards, built with INFRABEL as the end user. Multi-camera depth, on-device inference, feedback loop for event analysis.
Work with Velmeni on the data side of a dental diagnostics model: what the training set was missing, and how to generate the rest.
A multi-exposure fusion network that recovers lane markings when the camera is blinded.
Two weeks, fixed price, no obligation afterwards. You keep the report either way.