MobiCure VN - Multimodal Local LLM Chatbot for Colorectal Cancer: Patient Support and Clinical Training

Patients recovering from colorectal cancer surgery in Vietnam leave hospital with a stoma or a healing wound and very little continuous support. Reliable guidance is mostly in English, clinicians are hard to reach between appointments, and the questions that matter most: "Is this discharge normal?", "When is this an emergency?", are exactly the ones patients are least comfortable asking.

MobiCureVN is an iOS 26 application that puts a Vietnamese-speaking recovery assistant on the patient's ipad. Every component runs on-device: a 4-bit quantised large language model via MLX Swift, a local retrieval-augmented generation index built from 39 curated clinical documents (NCCN, NHS, UOAA; 1,238 chunks in SQLite with hybrid vector + full-text search), Apple's on-device Translation framework, and a vision-language model for wound-photo triage. No patient text, chat history, or wound photograph ever leaves the device.

Answers are grounded in the retrieved clinical sources and shown with citations. A dedicated emergency detector runs on the patient's original Vietnamese wording before any translation or generation step, so red-flag symptoms trigger a direct "seek care now" redirect rather than a generated answer. Input and output guardrails sit either side of the model to keep it an assistant, never a diagnostician.