92% in forty-eight hours

Inspection teams often have years of records and still cannot see risk spatially. High-risk assets look like healthy ones until something breaks, and then everyone is reactive.

At MECC 2025 (Aramco-sponsored), we had 48 hours and a five-person team. The product was not a paper accuracy number. It was a system someone could open and use while the clock ran.

We engineered features from 1000+ inspection records and trained a Random Forest that hit 92% accuracy on that set.

Label it correctly: hackathon / model result on the data we had, not a plant-floor SLA.

Why not a heavier neural net? Under a demo clock, a model you can retrain, debug, and explain beats one you cannot trust live. Classical ML was the engineering choice, not nostalgia.

I built the React Router dashboard (risk heatmaps, asset views, analytics) and integrated the FastAPI prediction service so inspectors were not bouncing between a notebook and a spreadsheet.

We added a hybrid fallback so the UI stayed usable if the model service dropped. A blank screen kills demos faster than a degraded score. “100% uptime” here means the prediction path was designed not to brick the dashboard, not that a refinery ran on our laptop.

We cut scope hard: heatmaps, assets, alerts, not a full CMMS. With a longer runway I would tighten evaluation splits, show inspectors clearer confidence, and stress outdoor/ops constraints earlier.

The portfolio story that matters: honest model selection, a service that fails soft, and a UI that makes risk spatial. Details live in the inspection case study.