MachinaCheck: Multi-Agent AI for CNC Manufacturability on AMD MI300X
MachinaCheck, a multi-agent AI solution, was developed by Syed Muhammad Sarmad and Sabari Doss R during the AMD Developer Hackathon in May 2026. This system streamlines CNC manufacturability assessments for small machine shops. Users can upload a STEP file along with material type, tolerance, and thread details, receiving a comprehensive report in just 30 seconds. Operating entirely on-premise with an AMD Instinct MI300X GPU featuring 192GB HBM3 VRAM, it ensures that proprietary STEP geometry remains within the shop. The architecture comprises a five-component pipeline involving a STEP file parser (cadquery, Python), operations classifier (Qwen 2.5 7B), tool matcher (pure Python database lookup), feasibility decision agent (Qwen 2.5 7B), and report generator (Qwen 2.5 7B), all orchestrated via FastAPI and LangChain, with vLLM for inference. The full pipeline executes in 25 to 40 seconds, and decision accuracy was validated on all GrabCAD test parts. Notably, LLMs are employed only for reasoning, while tool matching is deterministic. The developers suggest that Qwen 2.5 72B could enhance reasoning on the MI300X. The project is accessible on Hugging Face Spaces and GitHub.
Key facts
- Built by Syed Muhammad Sarmad and Sabari Doss R at the AMD Developer Hackathon, May 2026.
- Runs on AMD Instinct MI300X with 192GB HBM3 VRAM.
- Uses Qwen 2.5 7B, ROCm, vLLM, LangChain, cadquery, FastAPI, Next.js, Hugging Face Spaces.
- STEP file parsing via cadquery (OpenCASCADE) for exact geometry extraction.
- Tool matcher is pure Python database lookup, no LLM.
- Full pipeline end-to-end: 25 to 40 seconds.
- Decision accuracy correct on all test parts from GrabCAD.
- Zero bytes of STEP geometry transmitted externally.
Entities
Artists
- Syed Muhammad Sarmad
- Sabari Doss R
Institutions
- AMD
- Hugging Face
- GitHub
- GrabCAD
- OpenCASCADE