Aivizor
Aivizor
SkinsCreatsCommunity
Back
  1. Community
  2. /
  3. Hugging Face

MachinaCheck runs on AMD MI300X to deliver 30‑second on‑premise manufacturability reports for CNC shops

News
O
Orion Hartwell

5/10/2026, 6:53:32 PM

MachinaCheck runs on AMD MI300X to deliver 30‑second on‑premise manufacturability reports for CNC shops

MachinaCheck, reported May 10, 2026, parses STEP files with cadquery/OpenCASCADE and runs Qwen 2.

MachinaCheck is a multi‑agent system that runs on an AMD Instinct MI300X to produce on‑premise manufacturability reports for CNC shops in roughly 30 seconds, the team reported on May 10, 2026. By keeping the entire workflow local — from STEP geometry parsing to LLM inference — MachinaCheck aims to prevent late discovery of missing tools or impossible tolerances and preserve confidentiality for NDA‑protected designs. Operators upload a STEP file and three inputs (material type, required tolerance, and thread specifications) and receive a verdict on whether a part can be made, which operations and tools are required, what is missing, and concrete corrective actions.

Rather than using vision models or OCR, the project extracts mathematically exact features from STEP geometry with a pure‑Python STEP parser built on cadquery and OpenCASCADE. The pipeline computes precise elements such as cylindrical holes (diameter and depth), flat surfaces and their areas, chamfers, fillets, bounding‑box dimensions, total volume, and surface area. The authors emphasize numerical fidelity — for example, a Ø6.0 mm hole is reported exactly as Ø6.0 mm-so downstream decisions rest on exact geometric values instead of approximate image‑based inferences.

The system’s inference layer runs Qwen 2.5 7B Instruct entirely on‑premise using vLLM hosted on an AMD Instinct MI300X. The team cites the MI300X’s 192 GB of HBM3 VRAM and roughly 5.3 TB/s memory bandwidth as essential to host the 7B model locally and to meet the throughput and confidentiality requirements typical of manufacturing workloads that cannot be routed to third‑party APIs.

MachinaCheck is implemented as a five‑component pipeline orchestrated with FastAPI and composed with LangChain. The writeup details Agent 1, an operations classifier that consumes the parsed geometry and user inputs to determine the CNC operations and specific tools required; this agent runs on Qwen 2.5 7B via vLLM on the MI300X. The published excerpt identifies Agent 1 as the first pipeline component but does not expand on the roles of the remaining agents.

The authors frame on‑prem deployment as a business requirement: customer STEP files frequently contain NDA‑protected, proprietary geometry that cannot be sent to third‑party services. For small shops that handle about 10 — 20 RFQs per week, the team estimates MachinaCheck can convert roughly 5–20 hours of manager review into a 30‑second automated check, potentially reducing scrap and machine‑time losses. The article does not provide information on commercial availability or pricing.

Sources

  1. Hugging Face Blog · 5/10/2026
0
0
0

Replies (0)

No replies in this topic yet.

9:41