How WHOOP turned wearable data into an AI coach | WHOOP AI

Viviano Cantu is the AI Product Lead at WHOOP, but his path there did not start in product. He joined WHOOP in 2021 as a software developer, led the storefront team, and then in March 2023 moved over to start a new team around coaching. That same month, he and a few colleagues started ordering pizza on Friday afternoons and hacking on whether an LLM could actually coach someone throughout their day. Six months later, WHOOP Coach launched as the first generative AI coach in a wearable. He has led AI at WHOOP ever since, and along the way his team built AI Studio, an internal platform that lets anyone at WHOOP go from idea to a working AI agent in under ten minutes. In this episode of Product Talk, host and Planned Parenthood Direct Chief Product Officer Hannah Park sits down with Viviano to speak on how WHOOP turned wearable data into an AI coach, how they think about evals, synthetic personas, and the my memory feature, and what his engineering background actually changed when he crossed over into product.

Subscribe to the Product Talk podcast on Spotify and Apple Podcasts and catch every conversation with leading product executives. Episodes drop on Wednesdays.

Show Notes

1. Viviano did not start as a product person. He joined WHOOP as a software developer, led the storefront team, and stumbled into AI coaching the same way he has approached most problems in his career: by seeing something that should exist and building a quick prototype to find out if it could. The WHOOP Coach story began with a Chrome extension he built on the side that could summarize web pages, which he then pointed at WHOOP data, and the resulting demo was enough to get the executive team to stop everything and form a team around it.
2. His Boston Rat Czar side project is a useful window into how he thinks. He and his partner almost rented an apartment with a serious rat problem, and the 311 data that would have warned them was technically public but practically impossible to navigate. So he built a website that aggregates the data, maps it, and makes it searchable. The pigeon poop bridge in Seaport is apparently a known landmark on the map.
3. The Friday afternoon hack sessions that eventually became WHOOP Coach worked because they came with an implicit promise: whatever gets built here will not just die as a demo. People will show up on Friday afternoons and order pizza for a few weeks if they believe the work is going somewhere. The sessions ran for only a few weeks before the executive team told them to stop everything else and work on AI full time.
4. The team was building on GPT-3 when GPT-4 dropped mid-build. The difference was dramatic enough that they made the call to step back, delay the timeline, and rebuild the data querying engine to take full advantage of the new model’s capabilities. Viviano frames this not as a setback but as an example of holding a deadline loosely: set a date to create accountability, but know that the right shot on goal matters more than the original schedule.
5. The gap between a weekend AI health demo and a coaching product that actually works is the gap between impressing a single user with their own data and building something that reliably works for many people at scale, over years, without ever knowing which question is coming next. Viviano’s answer to that gap is evals. He considers evals the single most important thing anyone building an AI product needs to have solved, and most teams building health chatbots have not solved them.
6. Evals for coaching are not straightforward because coaching answers rarely have one definitively correct response. The approach WHOOP uses is to identify the dimensions of a wrong answer rather than the definition of a right one. If WHOOP Coach is building a workout plan, it can test whether the recommended duration matches the math, whether the intensity is appropriate given the member’s recovery score, whether it is recommending too much or too little. Getting specific about what wrong looks like is how you build meaningful evals when right is genuinely ambiguous.
7. Because WHOOP takes member privacy seriously, they cannot run evals on real member conversations. So they built synthetic personas: realistic fake member profiles with plausible data that can be used to test coaching quality at scale. The personas are designed to surface the kinds of edge cases and ambiguous situations that real members actually encounter, so the evals have real predictive value.
8. AI Studio is WHOOP’s internal platform for building AI agents, and the first version was built in about a week. It lets anyone at WHOOP go from an idea to a testable AI agent in under ten minutes. The key insight behind it was that prompts and evals change constantly, and keeping them in code meant product managers were essentially writing prompts and handing them to engineers to copy and paste. Moving it into an internal tool eliminated that bottleneck and dramatically increased velocity.
9. The evals-first principle that shapes how WHOOP builds agents: the prompt is not the most important thing. The most important thing is defining what you want the agent to do and how you will know if it is doing it. Once you have a solid eval, the prompt is just the variable you optimize to score well on it. The tooling exists to make that organic process fast even when you only know what right looks like when you see it.
10. AI Studio changed how engineers think about their work. Instead of building features, they now think about which parts of the app to connect to AI Studio, because each connection unlocks new things that product managers and domain experts can build without engineering support. The result is that AI is no longer something that lives inside a single team; it has become a capability that any team at WHOOP can reach for.
11. The privacy architecture behind WHOOP Coach was not the result of a deliberate upfront design decision so much as a natural outcome of building carefully. PII never leaves WHOOP, and the model receives physiology data but never identity data. All the friction in the system sits at the point where agents reach members, not at the point where people inside WHOOP prototype ideas. Make internal iteration fast, make member-facing shipping deliberate.
12. A new role is emerging at WHOOP that sits between engineer and PM, and it is becoming its own craft. The people who are best at it are either engineers with strong product taste and collaboration skills, or domain experts (fitness coaches, health specialists) who are passionate enough about deploying their knowledge at scale to learn the tooling. The job is essentially teaching the agent what right looks like, which requires subject matter depth more than it requires the ability to write code.
13. Entering the clinical space with EHR sync introduced a different kind of AI problem than WHOOP had solved before. Viviano uses the precision and recall framing to explain it: most consumer AI products optimize for recall, making sure the right answer is somewhere in the response. Clinical AI has to optimize for precision, making sure nothing in the response is wrong or harmful. The clearest illustration is that ChatGPT responds to “I have a cough” with a list of possibilities, while clinical AI responds with a clarifying question.
14. The My Memory feature, which lets WHOOP members see and edit everything the AI knows about them, was a deliberate choice to open up the system during a period when AI is still new enough that members reasonably want to understand how it works. It lets members correct mistakes before those mistakes compound through future coaching sessions, and it created an unexpected positive effect: members who saw sparse memory profiles were motivated to add more context, which made the coaching better.
15. There is also a practical privacy argument for My Memory: if there is something a member does not want the system to know, they can delete it. Viviano acknowledges that as models improve and memory becomes more embedded, the feature may need to evolve, but for now it gives members more control and builds more trust than hiding the system’s knowledge would.
16. The engineering habit that survived Viviano’s move into product is metric rigor. His entire team, including engineers, looks at product metrics daily: engagement, quality, latency, and cost. He is particularly focused on the relationship between those last three, because improving quality can silently degrade latency or inflate cost, and you can move one in the wrong direction without realizing it if you are not watching all of them together.
17. The engineering habit that did not survive is resource roadmapping. As an engineering leader, he could track exactly what each person was working on each week because the work was discrete and sequential. Product management does not work that way: a PM might be writing a PRD, aligning stakeholders, monitoring an experiment, and steering an AI agent all in the same week, and none of those map cleanly to a delivery timeline the way an engineering task does.
18. His thinking on AI-assisted coding discipline came from watching what happened when less experienced engineers used cursor and Claude Code without the underlying craft to guide them. Senior engineers could use these tools well because they knew all the intermediate steps needed to solve a problem, so they could catch when the AI was skipping something important. The principle he landed on: never just hand a task to AI and ship the output; stay in the loop at the right points in the process.
19. His practical AI workflow advice for product managers is three things: work out of one folder and drop all context into it rather than managing separate project folders; run many parallel threads at the same time rather than working sequentially; and use auto-approve with a sandbox so agents can run tasks to completion without requiring approval at every step. He runs about ten parallel tasks at a time in product work, which he acknowledges is lower than the number he ran as an engineer but still more than most PMs are doing.
20. In the lightning round, his lowest WHOOP recovery score was 1%, alcohol related. His favorite AI tool, the one he would fight to keep if he could only keep one, is Claude Code, including for product work, because agentic mode has changed how he works more than anything else. And the one habit his own WHOOP data actually got him to change: going to bed earlier, because fixing sleep makes everything else easier.