devagarwal2/cerebras-showcasw-demo

Bill of Vision is a demo project that turns a short video of any machine into a detailed bill of materials, the kind of structured parts list that factories, repair shops, and insurance claims are built around. Instead of a person watching the footage and typing up what they see, four separate AI agents work through the video automatically and hand back a finished, organized parts list within seconds. The process starts by pulling twelve evenly spaced frames out of the uploaded video using ffmpeg. From there, one agent identifies every visible part in those frames and reads any nameplate or label it can find, even in non English scripts, quoting the exact text it saw rather than guessing. A second agent works at the same time to estimate the machine's scale, dimensions, weight, and materials from the images. A third agent then matches each detected part name to a real entry in a parts encyclopedia website called bomwiki.com, which has over 176,000 known parts. A fourth agent stitches all of this together into a final hierarchical report, grouped by sub assembly. A core design goal here is trustworthiness rather than pure automation. The project takes deliberate steps to stop the AI model from inventing information it did not actually observe. Any part label the model reports must be something it genuinely read off the video, and any link to bomwiki.com must point to a real, verified entry rather than a guessed or made up web address. If no confident match is found, the system leaves that field blank rather than fabricating one. Progress is streamed to the browser live as each of the four agents finishes its work, so a user watching the page sees real status updates rather than a fake loading animation. On the technical side, the frontend is built with the Astro framework, and the AI reasoning is powered by Google's Gemma model running through the Cerebras inference service. This project was built as a demo for a hackathon or showcase event around the Cerebras platform.