Health condition flare detection agent

CATEGORIES

Go Deeper

I have a complicated chronic thyroid disease called Episodic Graves disease. This means I randomly will become hyperthyroid to the point I an hardly function until meds kick in. Because I have a complicated episodic version I have to go on and off meds as symptoms arise, even though the meds take 1-2 weeks to start working. I have an ML engineering background and gave claude code 9 years of apple health data and labels of my flares over the years with the goal of detecting these flares before they become obvious. I iterated on a pretty robust ML discovery project setup and harness and the results changed my life. Claude Code ran hundreds of experiments and engineered complex features that eventually led to a model that detected my validation data nearly 100% of the time 2-3 weeks early. I vibe coded a personal app I could check every day and within about a month of using I received my first early warning. I went in for a blood test the same day and saw early elevation of my T3 hormone and my doctor put me on a low does of meds before I felt a single symptom. For the first time my life was changed and I didn't have to experience the impending hyperthyroid flare. I shared this and went moderately viral on X, tik tok and reddit. 100s of people told me they need this for their unique condition. The problem, the model was an n=1 model specifically built for me. The solution, package my process with Claude Code into an agent that anyone could use to upload their Apple Health data and Labels to expect the full power of Opus 4.6 to discover their own personal ML model. The agent uses a complex ML discovery process to iterate on features and models until it hits its goal. The agent will try hundreds of complex features beyond what any health apps currently use(14d respitory rate Z score / 30 d zscore etc etc) and dozens of model param configs. This was essentially an early version of Andrej Karpathy's autoresearch for health ML specific research problems. I built a consumer app called Pharo Health where the users can connect apple health, create labels and send data to the agent in the cloud for processing. After a successful run the agent sends back a coreML model that runs 100% on device from that point on allowing the user to see daily predictions of their next flare. The next steps for this are just as exciting, a leading medical research hospital got wind of this and are interested in using the agent for more specific detection tasks for patients. This has led to early discussions on myself and the agent becoming the Technology partner on a wearable detection clinical study over the next year. Just an idea and a few weeks of vibe coding have changed my life and look to be leading to something much bigger. The age of personalized ML for our health is upon us!

Stack Used

Built with Claude Opus 4.5-6. App UI/UX Designed via a custom iterative process via Opus 4.6. App: SwiftUI, python fastAPI, docker compose Hosted: AWS Agent: Opus 4.6 agent in custom ML discovery harness(tools + prompts and injective prompts)