Daniel
De Brun

Building a benchmark for evaluating state-of-the-art multi-robot collaboration methods in NVIDIA Isaac Lab.

Led the re-platforming of an analytics platform serving 60 divisions and 3,500+ production lines onto a Databricks lakehouse: data refreshes every 5 minutes instead of hourly, at under one tenth of the monthly cost.

Owned the production GraphQL API through the migration, and led the micro-frontend platform that split a monolithic portal into six independently released applications.

Sole author of an agentic pipeline that maps supplier spreadsheets into one schema: deterministic lookup resolves 75% of columns before any LLM call, and nothing publishes without expert approval.

Shipped production OCR for stamped and engraved IDs on metal parts and built two MCP servers adopted across the company. Kept on past the initial term and moved to the corporate function with the core team.

Built multi-agent RL agents in JAX that adapt to a partner’s skill in Overcooked (CoGrid, JaxMARL); the project won the People’s Choice Award at CUCAI 2026.

Led experimental integration for the team’s paper submission on learned communication in multi-robot teams: implemented its baselines (MAPPO, IPPO and three learned-communication methods) and found bugs that would have invalidated results, including deliveries credited to the wrong order and a biased best-seed metric.

Founded the university’s autonomous scale-racing team in second year and led it for two years with 30 members each year. Designed the car’s autonomy software and mentored 15+ developers.

Moved ORB-SLAM3’s per-frame tracking work to CUDA: feature detection, descriptors, warp-level stereo matching and the image pyramid.

On an RTX 4080, feature extraction is ~10× faster (6.05 → 0.64 ms per image) and tracking time drops about 40% in stereo, monocular and RGB-D, with accuracy on EuRoC and TUM within ORB-SLAM3’s own run-to-run spread. A bit-exact mode reproduces all 8.0M CPU keypoints and every descriptor bit.

Predicts which parts will fail final test while they are still on the line, and decides whether pulling them for inspection pays: RAPIDS and XGBoost for training, TensorRT at the edge, Warp for simulation and cuOpt for the inspection plan.

On 1.18M Bosch production parts (0.58% failing), MCC 0.481 under the Kaggle competition’s protocol, near the silver-medal cut-off. Live features match training bit for bit on 354,824 parts.

A control plane for Claude Code and Codex research agents that stops them rerunning until a result looks good. The first run freezes the analysis plan, every distinct run counts against a trial budget, and only a person can score the held-out data.

A safety layer for running Interactive Brokers jobs unattended. Orders are dry-run and paper-only by default and live trading needs two explicit opt-ins; it enforces a kill switch and per-order, per-batch and daily limits, stays within IB’s rate limits, and records every order in an append-only ledger.

Autonomy software for the team’s scale race car on an NVIDIA Jetson. Camera, SLAM, TensorRT segmentation, control and actuation run as separate processes over lock-free shared memory; pure pursuit, LQR, MPC and learned controllers sit behind a safety supervisor that cuts throttle on any command older than 150 ms.