Behind the Development of WHOOP Coach

Our Chief Technology Officer, Jaime Waydo, shares more on how the team built a world-class performance coaching experience using the latest large language model AI.

In December of 2022, LLMs (LLM) made a huge dent in tech culture. For the first time, we were talking about Artificial Intelligence that could write papers, answer questions better than internet search engines, and even create poetry. For us at WHOOP, this got us thinking – how could we leverage the technology of a large language model to build on our mission of unlocking human performance? Could an LLM power our vision of building a world class performance coach?

Artificial Intelligence, or AI, is not new to WHOOP. The core WHOOP algorithms that calculate sleep, strain, and recovery are all machine learning algorithms. We use AI technology to take raw signals and turn them into an accurate breakdown of the different stages of sleep, detecting when you fell asleep and when you woke up, or when you worked out and what your heart rate was throughout the workout. It’s all done by machine learning.  

BUILDING WHOOP COACH

When LLMs made their splash last December, it was exciting. This was an interesting piece of technology that was more than just new, innovative technology – it had the potential to help our members and solve their problems in a real way. As we started playing with LLMs, the questions were, “What if you could talk to WHOOP 24/7?”, “What would that experience look like?”, “What would you do with a coach that had access to all of your data and could answer all of your questions about fitness, health, WHOOP, and more?”.

During the early days of WHOOP Coach in January, members of the engineering team who were interested in AI and LLMs began experimenting, brainstorming, and learning. Could they take their own data and pipe it into a large language model?  What questions would they ask that first prototype? We moved beyond just experimenting and formulated our problem statement: Build a 24/7 coach personalized to you and your WHOOP data. As we got into March and April, we brought in our design team and started playing with designs. What were the colors, the iconography, the animations? What was the entry point to your conversation? 

Another one of our key questions in the WHOOP Coach development process was the voice. This was the first time WHOOP would have a conversation with you. What tone does it use?  Does it use emojis? When to be playful vs. serious? We brought the team together to provide question prompts like, “If WHOOP Coach were a band, who would it be and why?” to help us identify the personality of the WHOOP Coach voice.  

HOW DOES WHOOP COACH WORK

The best way to understand how WHOOP Coach works is to follow how a member might ask a question like, “How did I sleep on June 10th of last year?”.

The first thing a member does is type “How did I sleep on June 10th of last year?” into the WHOOP Coach interface in their app, which is then sent to WHOOP servers. WHOOP then evaluates which model should answer the question. For specific questions like, “Where can I buy a WHOOP Hydroknit band?” we have an internal model that recognizes that question and returns an answer. For more general questions, the model removes any personal information from the question, replaces that personal information with a random identifier, and then sends the question to our large language model partner. For the question “How did I sleep on June 10th of last year?”, WHOOP determines that this is a general question that goes to the LLM.  

Next, the LLM will break the question down into segments such as topics and dates, including the topics that WHOOP Coach can cover, including sleep, strain, workouts, HRV, stress, and all of your favorite WHOOP topics. For the question “How did I sleep on June 10th of last year?”, the model determines that the topic is sleep – and that the date needed is June 10, 2022. Now, the LLM sends back to WHOOP that it needs sleep data from June 10, 2022.  

WHOOP has yet another model running that looks at the topic returned from the LLM, and then determines what related data is needed in order to give a well rounded answer. For the question “How did I sleep on June 10th of last year?” WHOOP will pull this member’s sleep data from June 10, 2022, along with any journal data and behavior impacts that may be relevant.

Yet another model searches across this data to find any correlations between the member question and WHOOP Performance Science data and research. All of this context is determined and sent to a WHOOP model that pulls out all personally identifiable information and sends that off to the LLM again  for the question “How did I sleep on June 10th of last year?”.

The LLM returns an answer, and another WHOOP model adds back in any of the personal data while making sure the answer is easy to read and understand. WHOOP also links to any relevant WHOOP articles, and to any areas of the WHOOP app. 

The best part of it all? This entire process takes less than 3 seconds for most questions.

THE FUTURE OF WHOOP COACH

WHOOP Coach combines the natural conversation delivered by LLMs and pairs that with several models of deep technology built by WHOOP to enhance the experience with data and analytics. As with all LLMs and AI in general, the technology continues to evolve at a rapid pace – and the team at WHOOP is continuing to stay up-to-date on how we can build the best possible experience for our members. Over the next few months, Coach will visualize your data, give your more insights on you and your data even if you don’t know what to ask, and answer more complex questions than ever before.

There is so much the WHOOP Coach will be able to do in the next few months – this is just the beginning of a world-class coaching experience that can help anyone and everyone achieve their health & fitness goals. 

Jaime Waydo is the Chief Technology Officer of WHOOP. She is a product and technology executive with over 20 years of experience leading across engineering, technology, product, operations (overseas & domestic), regulatory and safety functions. She has served a wide group of industry leading companies including Alphabet, Apple, Translucent Medical, Hansen Medical and NASA’s Jet Propulsion Laboratory.