An offline AI chat app for Android that runs a language model and Vietnamese speech recognition on-device, without a server, account or Internet permission.
Technical challenges
Small models on phones are slow and get dates and arithmetic wrong.
How I approached it
Calendar and arithmetic questions are answered by Dart code and given to the model as facts; prompts are capped and the system prompt state is cached. Time to first token fell from 8.3 s to 0.93 s, and the quality suite rose from 13/33 to 25/33.
Impact
All three phases delivered and tested on real Samsung devices; users install the model from a file inside the app.
Phase 1: product foundation
DonePhase 2: on-device LLM
DonePhase 3: offline Vietnamese voice input
DoneA configurable, weighted-probability prize wheel for events and giveaways — built solo, completed and released.
My personal Android assistant app — Tasks, Goals, Notes, and a spaced-repetition Learning engine, built by applying my reusable Android product foundation. Currently on hold while I finish the interview-question dataset the next Learning stage depends on.
Tell me about your product or challenge, and we can discuss how I can help.
Work With Me
