In the latest episode, Kyriakos Eleftheriou sat down with Roman Bugaev , CTO of Flo and Vlad Nedosekin Director of AI Platform, at the Terra API HQ in London, to discuss how they built the top health AI platform globally for women's health.

CHAPTERS

(0:00) Intro — Flo Health: From 20 People to 80 Million Users

(1:02) How Flo became the fastest-growing health company in the world

(1:48) Roman's early days: 20 employees, no revenue, product-market fit

(3:15) How did you know the product was a hit?

(3:25) The underserved women's health market — everyone was building Uber alternatives

(4:31) First ML: neural networks for cycle and symptom prediction

(5:31) Product evolution — from symptom tracking to AI-powered insights

(5:38) Building chatbots inspired by how doctors ask questions

(9:18) A/B testing at scale — Flo's custom experimentation platform

(11:43) Engineering structure: autonomous two-pizza teams

(13:51) Team mistakes — why separate mobile and backend teams failed

(15:29) Scaling from 4 servers to 600 services and petabytes of data

(17:53) "Whenever it's possible, we are NOT doing AI"

(23:15) Why temperature data is critical for ovulation prediction

(25:09) Why Flo is the most accurate period tracker — data diversity advantage

(28:04) Competition: "We don't really have real competitors"

(29:00) AI content creation — generating personalized medical articles

(31:01) Hallucinations vs. conflicting medical sources

(32:34) The three-person blind test: when AI disagrees with humans

(35:10) AI is more consistent than clinicians — but biased against women

(36:52) Fine-tuning open-source models on synthetic women's health data

(38:25) User profile: the foundation of Flo's personalization

(41:19) The digital avatar — your AI health twin that notices what you don't

(43:09) AI router: like a GP triage system for language models

(46:04) Router also controls tone of voice and remembers past conversations

(47:29) "Evaluation, evaluation, evaluation" — how Flo picks models

(48:40) Model stability: why proprietary model updates are dangerous for medical AI

(51:01) Anonymous mode: privacy that enables AI instead of blocking it

(53:49) On-device ML for the most sensitive health data

(56:03) Cloudflare outage — "when everyone is down, you're allowed to be down"

(56:58) Fine-tuning Llama 7B on Databricks — 10,000+ GPU hours per run

(58:07) Training vs. inference cost breakdown

(59:45) 100,000-token prompts: the hidden cost of medical AI

(1:01:05) Build vs. buy: "Build your competitive advantage, buy everything else"

(1:04:47) Value creation vs. value capture teams

(1:07:04) The future: AI that knows you better than you know yourself

(1:09:00) Time series models: the future of health prediction from wearables

(1:10:38) Q&A