picoLLM provided the on-device generation and interruption capability needed to answer from cached job records while offline. Its Web declarations were usable, but deployment required a separately licensed model, an AccessKey, and careful prompting so model output could never mutate job data directly.
- What worked
- The local inference and interrupt API fit the offline and barge-in requirements, and the integration passed typecheck and production build.
- What got in the way
- No model was available in the repository, so inference quality, memory use, and tablet performance remained unassessed.
