9+ years in product management. On this GitHub I explore the latest market and technology trends and turn them into AI project prototypes: everything here is personal work, built fully outside of and unrelated to my employment. I build at the intersection of spatial intelligence, autonomous systems and AI safety, ensuring agents perceive, reason, and act within strictly defined Operational Design Domains (ODD) in the real world.
Several of the projects below are in stealth mode: they live in private repositories until they are ready to ship. Public repos are linked where available.
My background in Physical AI goes back to 2016, when I worked with the iCub humanoid robot on visual recognition, semantic reasoning, and visual servoing systems. That early robotics experience shapes how I approach autonomous product development today.
SAR and Earth Observation AI pipelines for geospatial intelligence: satellite data acquisition, change detection, ship detection, damage assessment, and vision-language models for SAR imagery. Active, in stealth.
Real-time safety monitoring for autonomous fleets with teleoperation trigger detection. Adversarial scenario generation, ODD boundary testing, and applying UL 4600 and SOTIF (ISO 21448) standards to autonomous systems. Adversarial robustness tooling is public: pytorch-shield.
World foundation model evaluation for closing the sim to real gap: video prediction benchmarks, transfer metrics, and domain randomized synthetic data pipelines using NVIDIA Omniverse Replicator. Active, in stealth.
Deep RL foundations (Stanford XCS224R Deep Reinforcement Learning, completed), multi-agent reinforcement learning for coordinated autonomous operations using PettingZoo, Ray RLlib, and MAPPO, and deep RL for legged locomotion and manipulation. In stealth.
Ordered by current priority. Most projects are in stealth mode and developed in private repositories; public repos are linked.
| Project | Status | Repo | Focus |
|---|---|---|---|
| Operational Safety Monitor | Active | Private | Real-time safety monitoring dashboard for autonomous fleets. Teleoperation trigger detection built on Waymax and the Waymo Open Motion Dataset. |
| World Model Benchmark | Active | Private | World foundation model evaluation: synthetic data generation, sim to real transfer metrics, and video prediction benchmarks. |
| AV Safety Benchmark | Active | Private | AI safety evaluation framework for autonomous driving: adversarial scenario generation, ODD boundary testing, and safety-critical metrics. |
| SAR & EO AI Workflows | Active | Private | End-to-end SAR and Earth Observation pipelines: satellite data acquisition via the UP42 Python SDK, change detection, ship detection, damage assessment, and vision-language models for SAR imagery. |
| Adversarial Robustness Toolkit | Active | Public | PyTorch-based adversarial robustness toolkit for neural network defense. See pytorch-shield. |
| Deep RL Foundations | Completed | Private | Policy gradients, model based RL, and robot learning. Completed Stanford XCS224R Deep Reinforcement Learning coursework plus HuggingFace Deep RL implementations. |
| Synthetic Data Generation | In progress | Private | Domain randomized synthetic data pipelines for long tail edge case coverage using NVIDIA Omniverse Replicator and procedural scenario generation. |
| Multi-Agent RL Europe | In progress | Private | Multi-agent reinforcement learning for coordinated autonomous operations using PettingZoo, Ray RLlib, and MAPPO with shielded Deep RL. |
| Deep RL for Robotics | In progress | Private | Deep RL for legged robot locomotion and manipulation: terrain adaptation, contact rich tasks, and sim to real transfer in MuJoCo and Isaac Gym. |
I run an AI augmented PM Operating System across three environments:
| Context | Tooling | Use Case |
|---|---|---|
| Personal / AI Builder | Claude Code (personalized) | GitHub projects, research synthesis, personal productivity |
| Enterprise PM | M365 Copilot | Product strategy, roadmaps, business models, stakeholder communication |
| Production Engineering | Amazon Kiro | Requirements driven implementation tracking and status visibility |
| Layer | Technologies |
|---|---|
| ML / RL | PyTorch, JAX, Stable Baselines3, PettingZoo, Ray RLlib |
| Robotics & Simulation | ROS2, Gazebo, YARP, MuJoCo, Isaac Gym |
| Safety | UL 4600, SOTIF (ISO 21448), Waymax, adversarial robustness (PyTorch) |
| Synthetic Data & World Models | NVIDIA Omniverse Replicator, Waymo Open Motion Dataset, procedural generation |
| SAR / EO | UP42 Python SDK, Rasterio, GDAL, SentinelHub, vision-language models |
| AI Tooling | Claude Code, M365 Copilot, Amazon Kiro |