A computer-vision triage pipeline for structural inspection photos — the model architecture choice, a worked precision/recall tradeoff, the human-in-the-loop review workflow, and why this is a genuinely different model family from the text-focused LLMs covered earlier in this program.
The project brief: a structural inspection team generates thousands of bridge and building inspection photos every cycle, and reviewing every one at equal priority wastes a licensed inspector's limited time on photos that show nothing wrong. This project builds a computer-vision pipeline that flags photos likely to contain a crack, corrosion, or spalling — reordering the review queue by likelihood, never replacing the inspector's judgment about whether a defect actually exists. That scope boundary is deliberate and is argued for directly, not treated as a disclaimer.
The module walks the build the way a real AI engineer would: why this is squarely CNN/vision-transformer territory rather than a job for the language models covered in earlier modules, the dataset problem of scarce, expert-labeled, heavily imbalanced defect images, a worked confusion-matrix calculation showing why recall is deliberately protected over precision here, the confidence-ranked human review queue design, and the batch-versus-near-real-time deployment tradeoff. The full worked numbers, the complete tech-stack manifest, and the finished design reasoning are part of the unlocked module below.