A multi-agent AI system called MachinaCheck reduces CNC manufacturability analysis time from up to 60 minutes down to roughly 25-40 seconds, according to a project writeup on lablab.ai's Hugging Face blog. Built at an AMD Developer Hackathon by Syed Muhammad Sarmad and Sabari Doss R, the system runs locally on AMD Instinct MI300X hardware to keep proprietary CAD data off external servers, and combines deterministic CAD parsing with local open-weight language models to automate feasibility checks for machine shops.
Local Multi-Agent Architecture
MachinaCheck processes uploaded STEP files using a hybrid pipeline of Python tools and local LLM agents. Feature extraction is handled mathematically via the pure-Python cadquery library, producing exact measurements for cylindrical holes, surface areas, and bounding boxes without vision-model approximation.
The extracted geometry and user specifications—such as tolerances, thread requirements, and material types—pass through specialized workflow components:
- STEP Parser (Pure Python): Extracts geometric features in under 1 second.
- Operations Classifier (Qwen 2.5 7B): Determines required CNC operations and tooling based on material constraints and required machine precision.
- Tool Matcher (Pure Python): Uses deterministic database queries to cross-reference required tools against local shop inventory, avoiding LLM latency and hallucination risks.
- Feasibility & Report Agents (Qwen 2.5 7B): Synthesizes match results into structured decisions—flagging missing items like specific taps—and generates a manufacturability report.
On-Premise AMD Hardware for Privacy
Because machine shops often sign strict non-disclosure agreements covering customer CAD files, sending proprietary STEP geometry to public API endpoints creates confidentiality risk. MachinaCheck runs Qwen 2.5 7B entirely on-premise via vLLM and ROCm on an AMD Instinct MI300X GPU, which offers 192GB of HBM3 VRAM and 5.3 TB/s of memory bandwidth. Using a 0.5 GPU memory utilization setting (roughly 96GB VRAM), the project reports average agent inference latency under 3 seconds. No CAD data leaves the local shop network in this design.
Performance and Limitations
In testing with GrabCAD STEP files, feature extraction completed in under 1 second for parts with up to 50 features, and the full four-agent pipeline took between 25 and 40 seconds end-to-end.
The project's own writeup describes this as an initial hackathon prototype rather than a production-ready tool. It also notes that while prompt rules enforce structured JSON output from Qwen 2.5 7B, more complex industrial parts may require scaling to larger models such as Qwen 2.5 72B for reliable reasoning.