Mooonsync

Moonsync helps women achieve peak performance by connecting to your menstrual cycle data, giving personalized health recommendations, and taking action for you.

Comment

Inspiration

After 8 years in tech product management, Carla felt burnt out and exhausted. A series of symptoms dismissed by multiple physicians led her on multiple online rabbit holes searching for answers that eventually led her to move from San Francisco to the westside of LA to work on her fitness and spend a year learning about hormonal health and its impact on women’s performance. From there she understood the whole picture, where physical health, gut health, and work/energy optimization are all connected and constantly changing. Over the past year, Carla has conducted more than 100 user interviews, covering women of all ages. She discovered that 90% of women experience energy and skills pattern changes along their 28 day cycle and 93% are interested in understanding their cycles better.

Today there is no source of truth where women can go to get actionable answers on questions around their health, especially as it pertains to their menstrual cycle. 1 in 3 women have a chronic hormonal condition. Many of these women are looking for answers in a highly under-researched field or unaware that their issues are hormonal imbalances.

What it does

Moonsync is a multimodal women's health chatbot that uses body signals, medical research, and documented experiences to provide personalized answers to questions that would traditionally require hours of research from multiple sources. As we learn about each individual, we’re able to provide insights about specific behaviors and correlations to achieve peak performance. Women operate on a 28 day cycle, where each phase exhibits distinct characteristics around mood, nutrition, exercise, and skillset. Using Moonsync is like taking the best advice from your therapist, health coach, PT, and doctor at a given moment. We add the signals from your wearables as a hidden context to each prompt to provide proactive daily insights and take action for you.

How we built it

Tech stack We use Terra API (YCw21) to integrate biometric data, LlamaIndex to build our advanced RAG with Pinecone to host, Azure Open AI as our LLM of choice, back-end hosted on Modal, and front-end powered by Typescript. To power our picture ingestion we utilized OpenAI Vision. We use the Perplexity model to connect our LLM to the internet. We have also integrated Google calendar to schedule meetings for you.

Knowledge base Over 1,000 curated papers around menstrual cycle, energy levels, observational studies verified by 2 medical advisors (MD professors @ Stanford that are Zui’s mentors) Professional opinions from key doctors and psychotherapists in the space (blogs, youtube, videos, podcasts) over 700 documents Empirical evidence from other women: Tiktoks, instagram reels, and youtube videos (over 250+)

Key technical features LLM document summarization for academic papers LLM node classification Semantic splitting for more accurate chunking A multi-indexed query routing system connected to 5 knowledge agents (nutrition/diet, strongest skill set/feeling, exercise/fitness) to most accurately guide our LLM Auto Retrieval techniques to most accurately synthesize complete yet concise responses Multi-modal input ingestion via text and pictures.

Challenges we ran into

Recruiting users: We believe that this project will have a huge impact on 50% of the world population. We needed to find an initial test user to validate our concept. Earning someone’s trust in using their data to build our MVP, is not something we take lightly. Our goal is to always protect women’s data with the insights we might create. To our surprise, we had over 50+ responses from women who use wearables willing to supply their cycle data when reaching out to our network looking for an initial test user. We’ve had 12 women asking to pay to be a part of our pilot. We currently have a beta user waitlist of 260 women.

Thinking BIG: Carla had been working on the product concept and user validation for around a year and realized the idea was bigger than a task manager. GenAI developments in the last 3-6 months have made it possible for us to iterate fast, evolving the product to a fully fledged interactive knowledge base. Ideating this larger vision and validating it with users has been an exciting challenge.

Finding the right team: Having a highly technical product requires a highly technical team. Carla and Zui connected in a GenAI hackathon and found they had shared passion for the space. From there, they put together the team of 4. Zui and Carla “hired” (and fired) different developers looking for mission alignment and the right skillset. The current team is integrated by complementary skills: full stack development, GenAI eng, data scientist/healthcare/design, and product management/strategy/user research. The team plans on continuing to work together on it regardless of the outcome.

Building our RAG retrieval structure: Figuring out what's the best way to develop our RAG, structure indexes, and route queries to produce clear concise answers that combine biometric data with a women’s health context took some configuration. We tested out various query routing methods including auto-retrieval and multi-indexing. We knew that we wanted to structure our data into 4 agents (mood, nutrition, fitness, and general), and realized that maintaining these 4 categories as 4 separate indices in our vector database would make training easier. After completing our data ingestion process, we went back and restructured our vector database, allowing for fast and accurate retrieval. In addition to the right structure, we’ve also gone through multiple rounds of prompt engineering and user testing to make sure that beyond outputting correct information, our app’s responses feel natural, engaging and empathetic to users

Accomplishments that we're proud of

  • Creating the MVP in less than 2 weeks
  • Tech works --> ingested and classified over 1K nodes specific to our use case. Model works
  • Users love it --> we have a test user already implementing advice from our model and tracking results, and multiple users already leveraging it to get answers to their health questions.
  • Nothing like this exists.
  • Gave us a better answer than OpenAi and Gemini, Perplexity
  • Happy team
  • We were able to contribute improvements to the codebases of the tools we leveraged for our product: Llama index (Google calendar integration: timestamp passing fix + attendee list fix), Terra API (Dependency conflict fix + Continuous integration improvement), Modal app (Patch to fix modal volume deletion)

What we learned

  • Llama index has amazing documentation and is very developer friendly when building an advanced RAG
  • Different techniques about how to process user queries / query engines
  • Different ways to ingest and classify data smartly
  • Through our users, we learnt that we are solving a real problem

What's next for Mooonsync

In the future we see Moonsync supporting:

  1. Booking your doctor’s appointments, workout classes and get nutrition optimized meals
  2. Sending notifications to your loved ones to better support you
  3. Developing a licensed calendar extension to allow companies to offer this app as a benefit to their employees

Built With

  • azure-open-ai-as-our-llm-of-choice
  • azureopenai
  • back-end-hosted-on-modal
  • llamaindex
  • llamaindex-to-build-our-advanced-rag-with-pinecone-to-host
  • llm
  • modal
  • openaivision
  • perplexity
  • pinecone
  • rag
  • terra
  • terraapi
  • typescript
Share this project:

Updates