AI diabetes coach reveals ROI
Doug Warren published Part 2 of his ‘AI Diabetes Coach’ with a real-world architecture that automates wearable data ingestion and builds temporal knowledge graphs — he says the system hit 100% Time In Range during launch and raised his personal TIR from 82% to 98%. (x.com) Warren also claims the setup saves ~180 hours/year at an estimated $20/month total cost. (x.com)
Part 2 of Doug Warren’s AI Diabetes Coach series was published March 30, 2026 as “AI Diabetes Coach — System Architecture,” and it frames the post as Part 2 of 5 explaining the live system running on WarrenWeb. (blog.warrenweb.net) Warren splits the implementation into five core components named Calendar Service, Health Service, Review Service, Knowledge Graph, and an Integration Layer. (blog.warrenweb.net) The Calendar Service is described as automatically creating 13 interconnected Obsidian notes every morning at 7:00 a.m. and linking those notes into the Neo4j temporal graph. (blog.warrenweb.net) Warren says the Health Service pulls 288 glucose readings per day from the Dexcom API and also imports insulin doses and carbohydrate logs from Glooko, all via automated scripts. (blog.warrenweb.net) The Review Service packages seven days of context and sends it to Anthropic’s Claude for pattern analysis and a daily Plus–Minus–Next style coaching summary generated by the LLM. (blog.warrenweb.net) All patient data and metadata are stored in a Neo4j knowledge graph that Warren highlights for its temporal-query capability, including queries such as “show all days where exercise preceded a low within 12 hours.” (blog.warrenweb.net) An Integration Layer of Python services orchestrates the end-to-end flow between Dexcom/Glooko, Neo4j, Claude, and Obsidian, and Warren documents that his workflow reduced a manual diabetes review that previously took about 60 minutes down to an automated run that completes in under 10 seconds. (blog.warrenweb.net)
