Projects

GPU inference engines, custom CUDA kernels, and AI-agent systems I've built or am actively working on.

• 

  ai-hardware-engineer-roadmap

  Active

  My public learning roadmap and living resume — the path from AI inference to agent harness systems to hardware engineering, aimed at designing a physical AI chip. 200+ GitHub stars.

  • Roadmap
  • Physical AI
  • AI Chip

  GitHub
• 

  genie-ai-runtime

  Active

  A from-scratch AI inference engine for the NVIDIA Jetson Orin Nano, specialized for Qwen 3.5 4B and powering the GeniePod product — measurably faster than llama.cpp on the same hardware.

  • Jetson Orin Nano
  • Inference Engine
  • CUDA

  GitHub
• 

  hopper-qwen-72b

  Active

  Hand-written CUDA kernels serving Qwen 2.5 72B Instruct across 2×/4×/8× Hopper GPUs (TP=8), built to beat vLLM throughput — a ground-up study of multi-GPU inference at the kernel level.

  • Hopper
  • TP=8
  • Custom Kernels

  Private repository
• 

  triton-vm-prover

  Completed

  GPU-accelerated zk-STARK prover for Triton VM on the RTX PRO 6000 — 11,000+ lines of C++/CUDA, designed and built from scratch, reaching 10× the throughput of the top-ranked CPU prover.

  • CUDA
  • zk-STARK
  • C++

  GitHub
• 

  tether

  Active

  Deployment and inference optimization for Vision-Language-Action (VLA) models on NVIDIA Jetson — getting real-time robotics policies running on the edge.

  • VLA
  • Jetson
  • Robotics

  GitHub
• 

  genie-claw

  Active

  A tailored AI-agent harness for smart-home automation, tuned for low latency, tight context budgets, and higher task accuracy.

  • AI Agent
  • Smart Home
  • Low Latency

  GitHub
• 

  openclaw

  Active

  Maintainer of OpenClaw, a leading open-source AI agent framework — reviewing and shipping changes to the core.

  • AI Agents
  • Open Source
  • Maintainer

  Merged PRs

More on GitHub: @ai-hpc