Ekstraksi Data Formulir Downtime Tulisan Tangan Menggunakan Vision-Language Model Qwen2.5-VL Lokal pada Sistem Pencatatan Produksi PT Intiroda Makmur

Daily downtime records for the wire drawing process at PT Intiroda Makmur were written on paper forms and then retyped into a Visual Basic 6 desktop application available on a single office computer. Paper cannot be removed because occupational safety rules restrict phones near the machines. During an internship (5 July–14 August 2026), a web-based recording system was built on the company’s existing SQL Server databases, together with a module that extracts form data from photos using the Qwen2.5-VL 3B vision-language model running locally through Ollama, so no production data leaves the intranet. Extraction runs in two stages: a full-page photo for the header (Canny edge detection, perspective correction, CLAHE) and a close-up of the 48-cell downtime grid, rectified by a three-point affine transform. An ink-ratio gate skips empty cells before the model is called, prompts restrict the model to transcription, outputs are validated with Pydantic, machine codes are chosen only from the official master list, and operator names are matched with bidirectional string similarity (threshold 70%). A review screen keeps a human in the loop before data are saved. The ink-ratio gate reduced Section 2 processing time from 150–156 s to 2.0–33.6 s depending on the number of filled cells. All 45 unit tests (165 assertions) and 16 black-box scenarios passed; reading accuracy was evaluated qualitatively and still requires quantitative measurement.