Portfolio
Systems, research, and things that work.
A portfolio of difficult problems made inspectable—from agent architecture and cybersecurity research to the infrastructure beneath them.
Each case study follows the problem, constraints, approach, architecture, outcomes, and next questions.
1 project
| Project | Kind | Status | Role | Technologies |
|---|---|---|---|---|
| Transformer-augmented reinforcement learning in the Unity Food Collector environment | experiment | complete | Author of the PPO transformer-actor variant and the PPO baseline; architecture, training loop, stability fixes, experiment configuration, and report | Unity ML-Agents Food Collector (5×40×40 grid sensor, hybrid action space), Gymnasium wrapper over the Unity binary, PyTorch, Proximal Policy Optimization with GAE and clipped surrogate, CoordConv-ResNet visual encoder (31.1M parameters), Transformer encoder actor (d_model 512, 2 layers, 2 heads), Decoupled MLP critic on the raw observation, Cosine learning-rate schedule with 10k-step warmup, Aim experiment tracking |