I build interactive 3D physics and biological simulations, reinforcement learning locomotion labs, and serverless edge applications. Programming since the CP/M and MS-DOS days (C, Turbo C, Borland C++), combining three decades of systems engineering with modern AI-assisted workflows.
Zero-install 3D physics apartment and community locomotion hub where open-source bipedal robot ducks run real neural policies in the browser at 50 Hz, navigate obstacles and carpets, kick soccer balls, dock for charging, and hot-swap .ONNX skills in 1 click.
[1, 61] -> [1, 14]): BEST_alpha_walking.onnx, roulade.onnx (forward somersault & recovery), alpha_ground_pick.onnx, ball_kick_left/right.onnx, happy-hop (backlash-trained), single-leg balance, and SHA-256 verified kissqy R17-V167 6x Feetech HD-1910 suite + drag-and-drop custom .onnx benchmarking.quackd LLM mission orchestrator, DuckBuddy voice/vision daemon, duck-echo BPM rhythm & voice guard, duckmind pet lifecycle, quackback MCP coder assistant, and smart low-battery auto-docking.Interactive scientific platform simulating E. coli central carbon metabolism (genome-scale model iJO1366). Powered by Cloudflare Workers, Durable Objects, D1, WebAssembly (HiGHS LP solver), PixiJS, and Gemini AI.
Real-time 3D physics simulation (Symplectic Euler + XPBD + Coulomb friction) where articulated multi-legged walkers learn locomotion from scratch in the browser via PPO Actor-Critic and SEP-CMA-ES.
Dynamic Neuron Prosthesis and LECAP gate pruning (25%-50% sparsity) with low-rank residual compensation on Llama-3.2-1B (PyTorch / Hugging Face on RunPod GPUs).
gzip_static compression for MuJoCo & ONNX Runtime WebAssembly bundles (25.15 MB -> 6.48 MB), parallel ONNX/GLB asset prefetching, and kinematics fetch deduplication.
.ONNX policy testing hub (DuckSpace.org) for the open-source Microduck robotics ecosystem.