Custom mobile LLM with pattern flags and guided change flowsMobile-first LLM chain with RAG routing, tone controls & safety guards
Commissioned by health-tech startup EchoForma to turn a white-paper idea into a working conversational companion—CUUMI—that feels human, runs fast on mobile, and is safe by design.I trained a custom LLM in Google Colab and optimised it to run locally on modern mobile devices where possible, for speed and privacy.
The model was trained on anonymised journal entries blended with expert psychological guidance. It’s designed to notice patterns in what a user types—stress signals, triggers, unhelpful habits—and to flag them gently. From there, CUUMI guides the user through simple, structured flows (pause, ground, urge-surf, alternative soothing, plan the next action), always with safety, anonymity and clear boundaries as the priority.
CUUMI is the friendly face of the system. It’s lightly gamified—you “keep Cuumi healthy” by practising your own healthy habits—and it supports a community centred on care, encouragement and emotional strength.
Because core IP is under NDA, this page shares outcome-level detail only; any visuals are representations, not the production UI or model internals.