MediSentinelAI: A Safety-Critical Multi-Agent LLM Framework for Symptom Triage and Geospatial Healthcare Coordination

MediSentinelAI is an India-oriented progressive web application (PWA) that combines multimodal prescription ex- traction, symptom triage, medication reminders, sensory alerts, and geospatial emergency coordination. The author-reported implementation uses six role-specialized agents, an event-driven React client, service workers, IndexedDB, Gemini structured outputs, deterministic offline symptom rules, and Google Places with an OpenStreetMap fallback. In a reported 250-image evaluation, exact four-field record accuracy was 79.2%; a 180- vignette triage confusion matrix implies 91.7% accuracy and 45/45 observed critical-case recall; and Web Push succeeded in 488/500 trials (97.6%). These values supersede earlier summary percentages in the project dossier, which conflict with the supplied counts. The datasets, code, model prompts, and execution logs were not available for independent reproduction. Results are preliminary author-reported measurements and do not establish clinical effectiveness, regulatory exemption, or safe autonomous operation. Index Terms—Clinical decision support, emergency triage, geospatial health, medication adherence, multimodal large lan- guage models, progressive web applications