Biological-Age Agents at Thrive AI Health

Most of my work is clinical AI inside a hospital. For four months in 2025, I took the same toolkit somewhere very different: a consumer longevity app.

Thrive AI Health is the OpenAI Startup Fund–backed venture building a personalized AI health coach. I joined as an AI researcher from July to October 2025, in San Francisco, and worked on the layer that turns a person’s day-to-day health data into something the app can actually act on.

Biological age from wearables and biomarkers

Chronological age is the number of birthdays you’ve had. Biological age is an estimate of how old your body actually behaves — and unlike the calendar, it can move in response to sleep, activity, and metabolic health. That makes it a useful north star for a coaching app: a single, legible signal a person can watch change.

I built biological-age prediction models from wearable sensor and biomarker data — the continuous streams from a wrist device combined with lab-style biomarkers — so the app had a grounded estimate to reason about, rather than a generic risk score.

From signals to advice: LangGraph agents

A number on its own doesn’t change anyone’s behavior. The harder problem is turning it into specific, personalized guidance. I designed AI agents with LangGraph to analyze the wearable data and recommend interventions tailored to the individual, not the average.

1. Model. Estimate biological age and the underlying signals from the person’s wearable and biomarker data.
2. Reason. A LangGraph agent examines those signals in context — what changed, what’s trending, what’s actionable.
3. Recommend. The agent proposes personalized interventions — the concrete next step, not a generic tip.
4. Trace. I debugged the agents’ interactions with LangSmith, tracing each run to see where reasoning went right or wrong.

  Wearables + biomarkers  ──(via Junction)──►  data pipeline
           │
           ▼
  Biological-age model   ──►  physiological age + signals
           │
           ▼
  LangGraph agent  ──►  reasons over the signals  ──►  personalized intervention
           │                  (traced in LangSmith)
           ▼
  Delivered in the app   (iOS + Android)

Getting the data in: Junction

Wearable data is messy and comes from many devices. I integrated wearable data from Junction — a health-data aggregation platform that normalizes data from hundreds of devices and lab tests — into the AI analysis pipelines, so the models and agents worked from one clean, consistent stream instead of a dozen bespoke device formats.

Shipping it: iOS and Android

This wasn’t a notebook experiment. I deployed Python code to the iOS and Android app stores, so the modeling and agent work ran inside a real mobile product that real people used.

The throughline

Thrive looks like a departure from my hospital work, but it’s the same core skill: put an agent on top of messy health data and make it do something useful and personal. The stack — LangGraph for orchestration, LangSmith for observability, careful handling of sensitive health data — is exactly what carries back into clinical settings, where the stakes for privacy and reliability are higher still.

This describes my own contribution as an AI researcher at Thrive AI Health; the product is Thrive’s. Nothing here is medical advice, and no proprietary details are disclosed. Biological-age estimates are research-grade signals, not clinical diagnoses.