About this project
Most defect detection tools stop at "crack detected." That's not useful on a factory floor. You still have to figure out why it cracked, what caused it, whether it's urgent, and what to do about it — and you're doing all of that manually. CruxAI is our attempt to fix that. You upload a photo of a component. The system finds the defect, draws a box around exactly where it is, generates a heatmap showing what the model was looking at, and then — this is the part we're proud of — it actually tries to explain what went wrong. Not just a label. It pulls from a vector database of maintenance manuals and past failure records, reasons through the findings using a chain of AI agents, and gives you a root cause (e.g. "material fatigue from uneven cooling during heat treatment") along with specific next steps. Before any report goes to the engineer, a Verifier Agent checks every claim against the retrieved evidence. If something's not grounded, it gets flagged. Nothing leaves the pipeline unaudited. The whole thing is packaged as an MCP server on NitroStack — so any AI agent or studio can call into it, trigger inspections, search the knowledge base, or pull machine health history. We also built a full dashboard in React where engineers can upload images, review findings side-by-side with heatmaps, approve or revise reports, and export them as PDFs. We support three different vision tasks — anomaly detection (MVTec), surface defect classification (NEU, 99.7% accuracy), and steel segmentation (Severstal) — so it works across different line types and materials. The goal was simple: give the engineer the answer, not just the alert.
Manufacturing & Industry 4.0 track
Create intelligent systems for smart factories, predictive maintenance, quality control, and supply chain optimization.