Maintenance work is often documented poorly or not at all — technicians are focused on the task, not on writing it up afterwards, and structured documentation is time-consuming to produce manually. Yet this documentation is valuable: it supports training, quality assurance, and knowledge transfer between technicians. A prior Demola team showed that feeding a 40-minute maintenance video into a large cloud-based model with a minimal prompt could produce surprisingly detailed, tool-level documentation automatically. The open question is whether this same capability can be achieved with smaller, locally runnable models — which would make the approach viable in privacy-sensitive or connectivity-constrained field settings.

Key questions to be answered in the project:
1. How accurately can structured maintenance documentation (steps, tools, materials, sequence, timing) be extracted automatically from video?
2. How much prompt engineering is needed to get reliable structured output from models of different sizes?
3. At what point does a small, locally runnable model become 'good enough' for this use case compared to large cloud-based models?
4. What are the privacy implications of video-based documentation, particularly around recording workers and GDPR compliance?
5. How should the extracted documentation be validated or corrected by a human before being used for training or quality purposes?

In this project we aim to...
- Build a working pipeline prototype that extracts structured maintenance documentation from task videos.
- Compare performance across at least two model sizes, including at least one small, locally runnable model.
- Document the privacy and deployment constraints relevant to using this approach in real maintenance environments.