BRKMYR/AI Builder Portfolio
Each project starts from a customer problem, ships a working system, and reports the numbers that say whether it works. Written the way a product manager writes: problem, solution, success criteria, KPIs.
Projects
Safe Autonomy · Release Assurance
Benchmarks produce scores. A release needs a decision. Assurance Gate checks eval results against thresholds hashed and published before the run, reports every rate as a Clopper Pearson bound, and links every claim to the run behind it.
Live dashboard · NO GO on demo data · 2 tracks
Autonomy · Fleet Ops
Robotaxi teleoperators can't watch twenty screens. This dashboard ranks vehicles by risk, surfaces the specific reason a car needs attention, and lets the operator rewind.
Video · 3 scenarios · 7 trigger kinds
Defense · MUM-T C2
One human, several unmanned wingmen, time pressure. INTENT is the operator console for tasking, engagement authorization, and AI decision logs that record every call the system made and why, sitting on top of a MARL testbed.
Video · 3 scenarios · AI Decision Logs
Autonomy · Benchmark
Every AV company claims their planner is safer. None of the claims are comparable. This is a vendor-neutral scorecard: 60 scenarios, 4 families, one composite score.
Leaderboard · 180 runs · 3 baselines
Frontier AI · Evals
"Looks realistic" and "is a good world model" are different things. Six metrics, six failure modes, closed-form baselines so the harness itself can be validated.
Leaderboard · Toy suite · v0.2
Earth Observation · SAR
Radar imagery works at night and through cloud, but the tooling is scattered across research code. Five notebooks take Sentinel-1 tiles all the way to a shareable GeoJSON intelligence product.
5 notebooks · Sentinel-1 · Free data
Earth Observation · VLM
Ask "how many vessels are in this port?" of a radar satellite scene. Get a number and a confidence trace. Refuses questions the imaging physics can't answer, so no hallucinated counts.
Live console mock · Real API responses
Robotics · Master's thesis · 2016
Master's thesis on the iCub humanoid at TUM. A CaffeNet CNN with attribute and affordance classifiers and a KnowRob ontology on top, so the robot recognised objects it was never trained on: 0% stand-alone, 65% with semantic reasoning.
Thesis · Caffe · KnowRob · 0 to 65% on unseen objects