Traditional order estimation in CNC machine shops is a notorious productivity sinkhole, consuming up to 20 hours of a senior engineer's week. According to the developers of MachinaCheck, manually auditing a single technical drawing can take an hour of meticulous labor, requiring the engineer to cross-reference dimensions, tolerances, and tool availability. In this industry, a single misread parameter is a costly mistake that leads to scrapped workpieces and downtime for expensive machinery.
MachinaCheck, a project unveiled at an AMD developer hackathon, offers a radical shift toward automation. By translating the physics of metallurgy and part geometry into a multi-agent pipeline, the system slashes analysis time from an hour to just 30 seconds.
The architecture of MachinaCheck represents a deliberate pivot away from the cloud-centric mainstream in favor of on-premise hardware powered by AMD Instinct MI300X accelerators. As lead developer Syed Muhammad Sarmad points out, third-party APIs from the likes of OpenAI or Anthropic present unacceptable reputational and legal risks for the industrial sector. STEP files for aerospace or medical components are multi-million dollar assets; no shop is willing to risk leaking them to the cloud where they might be used to train someone else’s models. With 192GB of HBM3 memory and 5.3 TB/s bandwidth, the MI300X allows the team to run a Qwen 2.5 7B Instruct model entirely within the factory’s secure perimeter via vLLM, delivering privacy at the infrastructure level rather than just on paper.
To solve the persistent problem of AI hallucinations, the developers implemented a hybrid scheme combining deterministic logic with LLM agents. First, a Python-based parser using the cadquery library extracts mathematically precise geometric data, such as hole diameters and chamfers. This eliminates the "imagination" issues common in standard computer vision models. Then, a Qwen 2.5-based agent cross-references these figures with steel grades and tolerances to determine if the equipment can handle the load. A final agent checks the requirements against real-time warehouse inventory to provide a definitive verdict on whether to accept the job.
The shift to a GPU-centric model on the shop floor isn't about following trends—it’s a pragmatic calculation of the cost per drawing in tokens versus the hourly rate of a highly skilled specialist. The days of viewing AI as a mere chatbot toy are over. We are entering the era of cold inference economics, where local AMD hardware becomes a legitimate tool for hedging manufacturing risks.