we research it. we build it. we sell it.
Latent Boom is an R&D lab. Consulting pays for research. Research drives products.Everyone here holds or is doing a PhD. Everyone is building something for the future.
Three cities, one lab.
Research and hardware in Mons. Recruiting and academic work in Mostaganem. Clients and commercial work in New York.
Three fields, one loop.
Research
We pick problems that are still open and work them for years, not sprints. Methods get published and the code gets released.
Consulting
A handful of engagements a year fund the lab. Data, models, edge deployment. The client gets a system that ships; we get runway.
Ventures
When a result holds outside the lab, we spin it into a company and staff it.
Synthetic Data Generation
Generating the data you cannot collect. Images and video, but also text, tabular records, time series and raw sensor traces. We build the set, measure whether it is balanced, and only then spend money on labelling.
Edge AI / Embedded Systems
Models that run on the device, in the weather, on the power budget it actually has. Vision, audio and sensor streams, quantised and pruned so they keep working after the van drives away.
Stereo 3D Perception
Stereo, lidar and point clouds turned into distance, volume and clearance. Built for sites where a single flat frame cannot tell you whether something is about to go wrong.
Industrial Applications
The events you need to detect are the ones that almost never happen, so the recordings do not exist. We generate them, across camera feeds and sensor logs alike, then prove the detector holds on the real site.
Autonomous Systems
Perception when the sensor is losing: direct glare, night rain, spray, occlusion. Fusion across cameras, lidar and radar so the road is still there when one of them goes blind.
LLMs in Production
Fine-tuning open models on your own corpus, in your own language, then building the harness that proves it got better. Chat assistants that answer from your documents, with the retrieval and the refusals tested rather than hoped for.
Agentic Systems
Models that call tools, read a codebase, label a dataset, or drive an internal process end to end. We build the orchestration, then instrument it so a failure is visible instead of silent.
Data Management Systems
Most teams cannot say what is in their training set. We build the catalogue, the versioning and the labelling pipeline that makes a corpus something you can audit, split and trust, whether it holds images, documents or sensor logs.
Grasping and navigation where no dataset described the room.
Recognition and diarisation in noise, on device.
Small models that survive the drift after deployment.
Unity3D twins built before a camera is installed.
Ventures
All ventures →Hawiya
A national digital authentication platform. Verify once, reuse across public and private services. The model itsme established, built for markets without it.
Minelead
Lead generation at scale: production crawlers, a company and contact graph, thousands of monthly subscribers. Co-founded and run from 2019 to 2021.
We love hard problems.
We take a small number of engagements each year. If it needs new research, it is the right fit.