Here I walk through each problem and its constraints, why I tackled it the way I did and where it led me. If you're facing something similar, tell me about it.
5 projects · timeline
To understand a space, a robot needs to separate every object and know what it is. But those boundaries aren't always clear. Three improvements give more consistent instances and a fusion five times faster.
A colonoscopy SLAM has to recognise the places it has already been through. Bag of Words was replaced by a deep-learning network ported to C++: it performs on par and proposes more consistent candidates.
Where was an old photo taken from? A Structure from Motion pipeline written from scratch reconstructs the place in 3D and locates the photo within it, less than a degree away from COLMAP.
From scattered spreadsheets and notebooks to a single platform. Client onboarding, site visits, certificates and invoicing are automated, with AI sorting the paperwork and human review at every step.
From one figure every two months to every water meter, hour by hour. An IoT platform for villages and hotels: readings over LoRaWAN or M-Bus, from several manufacturers, in a single digital model with alerts as soon as something goes wrong.
