AI Diary Generation & Edge SLM Research

An R&D track exploring how much language capability can be pushed onto a phone. The product goal was a diary that writes itself: the app reads a day's images and location trail and generates an itinerary and narrative entry from them.

The research half was model selection and conversion — evaluating small language models from Ollama and HuggingFace against a mobile inference budget, then converting Vision Transformer models for Android deployment with ai-edge-torch and ExecuTorch. On the data side I used KMeans clustering over image embeddings to group a day's photos into coherent episodes before any text generation happened, which sharply improved the quality of what the model had to work with.

The serving half was a FastAPI backend that exposed the LLM endpoints and let us prototype the product experience while the on-device path was still being optimised.