Denis ZeniosA startup needed to see drones in the sky reliably, tell them apart from the birds and planes that fool most detectors, and know where each one was heading. I built a real-time system that does all three, and beats the published state of the art at drone detection on every object size, on the same public benchmark. On one brutally hard clip the best published model missed the drone in all 300 frames; mine found it in 262.
Read case study 4 minMachine learning that runs on real robots, on edge hardware, in real time.
Is the robot too slow to react? Does it miss what's right in front of it? Email me what's failing and the hardware it runs on, and we discuss what needs to be done.
Case Studies
A closer look at how a hard problem became a working system.
A vision-guided robot kept losing the person it was meant to follow whenever they moved too quickly. The problem was diagnosed as a concurrency-model failure, and the system was rearchitected from Python to Rust, after which the robot held its target where it had previously lost it, with end-to-end latency roughly halved.
Read case study 3 minWhy does a vision-language model (VLM) that can tell a kitchen from a hallway still drive a robot into a wall when asked to grab a coffee? It knows what it wants but has no sense of where the walls are. The fix was to stop letting it steer: the robot remembers where things are, and the model only names what it wants.
Read case study 3 minOffers
Fixed price, fixed timeline. Email me and let me know which one fits.
It sees but reacts too late.
Edge latency audit
A written analysis of your full codebase: why the robot is too slow, and what to do about it, ranked by impact.
It can't find what matters.
Perception-failure diagnosis
A written analysis of your perception pipeline: the technologies in use, a classification of what is failing, and recommended fixes with estimated improvements.
Something else is failing.
Set up a 30-minute call
A conversation about what is failing, where the problem most likely sits, and which next step fits.
What clients say
From the reference letters of clients and collaborators.
Much of the difficulty was in the robot rather than the code. Its SDK was poorly documented, and a fair amount of the machine's behavior had to be determined by running it and observing it rather than by reading what it was meant to do. He worked through that independently. Midway through, I saw the robot hit an obstacle and tip over, and asked what could be done. He suggested using the lidar already on the robot to stop it acting close to an obstacle, and implemented it within the week without the rest of the schedule slipping.
He began by reproducing the published model we were benchmarking against, matching its reported results to within about one percent, before building a new model that outperformed it. He worked largely independently. I gave him a broad problem, and he returned a scoped one, solved and well documented. The work he handed back was clear and checkable. I would be glad to work with him again.
He is the kind of engineer who can take a loosely defined problem and turn it into a working system. The code he leaves behind is tested, documented, and structured so that others in the lab can build on it.