QA Vision
Paint-quality control on car bodies, from a photograph
Problem
Paint inspection on the line comes down to an inspector's eye under a lamp. One is stricter than the next, attention drops by the end of a shift, and there is no way to settle a dispute: no frame was kept.
Solution
An operator uploads a photo of a part, computer vision finds paint defects and boxes them, and every finding lands in the detection registry. The plant manager sees shift KPIs and defect trends and exports to Power BI or Excel — with no hand-filled inspection logs.
Engineering decisions
- The model is a replaceable part: Roboflow, a local YOLO and an arbitrary HTTP API all implement one interface — load / detect / is_available. Moving to the local model is a config edit, not a dashboard rewrite: inside a plant the internet may exist only in the office.
- Severity is a transparent formula, not a second network: defect area and confidence, three thresholds. You can explain it to a process engineer.
- The registry is separate from the interface: detections go into daily CSVs with a fixed schema and KPIs are computed on top, so the same data opens identically in the dashboard, in Power BI and in Excel.
What went wrong
- We didn't train our own model from scratch. In three days the choice was between one model and no product, or the whole pipeline with a pluggable detector — we took the second.
- An unavailable module returns "clean". If the detector fails to start, the frame is marked CLEAN and goes into the registry. On a showcase that passes unnoticed; on a line it is a silent miss.
- There are deliberately no accuracy figures on show: we never assembled a validation set worth trusting, and printing a flattering percentage next to a plant's logo is exactly the case where a number is worse than no number.
Result
A working pipeline from operator login to a file in Power BI: three independent services, a pluggable detector, a detection registry and roles scoped per plant. We didn't place — three other teams did, and that's a fair outcome: in three days we produced a pipeline, not a shop-floor-ready model.
What I'd do differently
First thing to redo: a model failure must raise an error, not return a "clean" verdict. Second, a file-based registry won't survive several lines writing at once; the path to PostgreSQL is written up in the project but wasn't walked in three days.
How it went
Qostanai AI-Sana Industry Hackathon, 1–3 November 2025, at Dulatov University. Organised by the International Telecommunication Union, Qostanai Hub and the university; the case came from the car maker Allur. Two days of building, a third for the pitch. A prize fund of 1m tenge and 13 teams in the final: university and college students, and school students. Mentors came from NIT, the video-analytics teams of Wildberries & Russ, the startups Aman Online and Impro, ITU experts and Allur's own engineer on the case.
Certificate
Signed by the president of Dulatov University, the ITU regional director for the CIS, the head of Allur's corporate university and the director of Qostanai Hub.