Vision-Language Models for Automated Extraction and Anomaly Detection of Handwritten Field Data: A Lightweight Industrial Workflow

Manual transcription of handwritten tabular data from field sheets into corporate information systemsis a pervasive source of latency and human error in industrial operations. This report describes andregisters a lightweight workflow in which a Vision-Language Model (VLM) replaces manualtranscription: the operator photographs the handwritten sheet, submits the image to a VLM via a fixedprompt, and receives a structured, downloadable spreadsheet with all values extracted andanomalies flagged — in under 30 seconds. Structured output is enforced through Pydantic v2schema validation combined with Anthropic's tool_choice mechanism, eliminating free-textresponses. Anomaly detection is performed statistically (±2.5σ over the batch) rather than by fixedthresholds, making the workflow applicable to any measurement domain without reconfiguration. Thesolution was cross-validated against two independent VLMs with consistent results. For a network of14 industrial facilities running two daily sampling sessions, the estimated annual saving exceeds1,750 operator-hours, at an API operating cost below $420/year across the entire network. Humanoversight is preserved as a structural requirement: the system detects; the operator decides.