LLM-assisted workflow generation for low-cost visual inspection and image analysis

This work presents a lightweight framework that uses large language models (LLMs) to co-generate user-interface schemas (JSON) and executable Python modules for custom image-analysis pipelines. Modules follow a simple, fixed interface, and the active configuration can be exported as a JSON. We evaluate module generation across three operator classes and four LLMs (48 generations): first-run success was 66.77.9 with typical fixes limited to imports or signature mismatches. Feasibility is demonstrated on plywood-knot counting: with various illumination and zoom the exported pipeline processed most test images correctly, however with larger changes in lighting/scale parameter retuning might be necessary. The approach targets low-cost settings by relying on open-source libraries and on possible deployments on low-cost embedded platforms (e.g., Raspberry Pi or Jetson) is feasible subject to processing-throughput constraints. The framework reduces integration effort for users with basic Python skills and supports rapid iteration in domains such as materials microscopy and visual inspection.