Product strategist and builder in B2B deep tech: data & AI platforms for autonomous vehicles, intelligent machines and systems.
Close to a decade in product management. Here I turn where the market and the technology are moving into working prototypes, built like products: a problem worth solving, a system that runs, success criteria set before the result, and evals that say whether it works. Personal work, unrelated to my employment.
Most repositories are private; each project is presented as a case study in the AI Builder Portfolio.
| Project | What it does | |
|---|---|---|
| Kartograf | GitHub | Ask a city map a question. A local language model writes spatial SQL over open map data, shows the answer on a map with the SQL behind it, and a 20-question eval says how often it is right. |
| Waymax Safety Monitor | GitHub | Which vehicle needs a human now, and why. A risk-ranked dashboard for remote operators of driverless fleets. |
| INTENT Operator Console | Case study | One human stays accountable for several machines, with authorization nothing can bypass and a log of every decision. |
| Label Quality Engine | Case study | Two sensors agree, or a human looks. Makes the line between machine labels and human review a priced decision. |
| World Model Benchmark | Case study | Six metrics against six failure modes, with baselines that validate the harness itself. |
| AV Safety Benchmark | GitHub | A vendor-neutral safety scorecard for driving planners: 60 scenarios, one composite score. |
| Assurance Gate | GitHub | Eval scores in, release decision out. Thresholds are fixed before the run and every claim links to its evidence. |
| SAR Intelligence & VLM | Case study | Ask a radar image a question, and get a refusal where the physics cannot answer. |
| iCub Object Recognition | GitHub | Master's thesis: a humanoid robot names objects it has never seen. |
| Layer | Technologies |
|---|---|
| Evals & Release Gates | pydantic, numpy, scipy, Inspect, Clopper Pearson bounds, UL 4600, SOTIF (ISO 21448) |
| Autonomy & Simulation | Waymax (JAX), Waymo Open Motion Dataset, PettingZoo, Gymnasium |
| Perception & Data Engines | PyTorch, CLIP, Grounding DINO, OWLv2, PandaSet, CVAT |
| Agents & Geospatial | Ollama (Qwen3, Llama 3.1), DuckDB spatial, Overture Maps, kepler.gl |
| Earth Observation | Rasterio, GeoPandas, Sentinel-1 |
| Serving | FastAPI, static HTML on Cloudflare, Hugging Face Spaces |
AI is moving into the physical world, and the key enablers are platforms: the data machines learn from, the evals that check them, and the release that puts them to work.