Vision QC (Fiber Installation Audit)
Vision LLM system that checks fiber installation photos against an ISP's photo standards. The proof of concept won the tender, and production is underway.
From proof of concept to production
The proof of concept put open and closed vision models behind one provider layer and compared them: Qwen3-VL on a self-hosted vLLM server; Claude, Kimi K2.5, and Amazon Nova on AWS Bedrock; and Gemini on Vertex AI. I chose a self-hosted Qwen3-VL-32B to meet data residency requirements, and added an OpenCV quality gate that turns away unusable photos before any model call.
The proof of concept won the tender. Maknadata is now the vendor, and the production system is built so its results can be trusted and its standards can change safely.
How it decides
The model reports what it sees in each photo, point by point, with its evidence. The verdict comes from versioned rules rather than from the model, so the same observations always lead to the same result, and every result can be traced back to what was seen.
What I built
The experiment workflow the AI team uses to tune rules, prompts, and models against ground truth, with the cost of every run, and the assessment service that runs the checks through an API and keeps a record of every prompt, model, and call.