HerHealth

Women’s symptoms often get missed because medicine studied men. HerHealth is an AI teacher that helps women recognize the signs, with every fact checked against trusted medical sources.

Comment

Inspiration

For decades, medical research studied mostly men, and the "textbook" signs of many conditions still reflect that. Women's symptoms often look different, so they get missed or dismissed by the people around them, by clinicians, and by women themselves. Chest pressure gets called anxiety, pain gets called normal, exhaustion gets called stress.

The information to recognize these signs exists in trusted medical sources, but it's scattered, dense, and rarely framed around women. We wanted a way for anyone to learn it quickly, and to trust what they learn.

HerHealth is for women whose symptoms get brushed off, the people who care about them, and the medical students who will be their clinicians next.

What it does

HerHealth is an AI teacher, a talking 3D avatar, who turns any health topic into a short, interactive lesson about how it affects women.

  • Make a lesson on anything. Type a topic or paste a link and get a full lesson in about 30 seconds. Every lesson centers on women: how symptoms can differ, pregnancy, periods, menopause, and differences in risk. If the trusted sources describe no difference for women, the lesson says so instead of inventing one.
  • Learn by doing. Each lesson starts with a quick true/false check of what you believe. Then you meet a woman in a realistic situation whose symptoms are easy to dismiss, and you decide what's going on. The explanation maps her symptoms onto the body next to what people usually expect. The same check at the end shows what changed.
  • A teacher you can talk to. She speaks with a natural voice and lip sync. You can ask her anything by mic or by typing, and she answers in a sentence or two from the lesson's sources. If you describe an emergency, she tells you to call 911.
  • Everyday or med student level. Med students get a clinical vignette, a board-style question, red flags, and what else to consider.
  • Chrome extension. Turns the page you're reading, or a highlighted phrase, into a lesson.
  • Ready-made lessons on heart attack, endometriosis, and ADHD, three areas where women's signs are often missed.
  • Impact page. Shows how learners' scores move from the first check to the last.

Every fact links to its source, with the exact sentence it came from.

How we built it

  • App: Next.js 16, TypeScript, and Tailwind on Vercel. The whole lesson floats in CSS 3D space and leans with your pointer.
  • 3D teacher: React Three Fiber renders the avatar. Her mouth follows the real audio through a Web Audio analyser, so she only moves when she speaks.
  • Voice: Cartesia Sonic. Audio is cached by a hash of the text, so repeat lines are free.
  • Retrieval: Elastic Serverless holds the trusted passages for the built-in lessons. Semantic search finds the ones that matter for each case.
  • Tutor agent: An Elastic Agent Builder agent, the HerHealth tutor, answers learners' questions during a lesson. It calls an ES|QL search tool over the index and answers only from what that tool returns, with Claude as a fallback.
  • Any-topic sources: The MedlinePlus API finds trusted passages for any topic in Make a lesson, including extra passages about the topic in women, pregnancy, and menopause.
  • Lesson generation: Claude writes each lesson as structured output, reviews it against the sources, and answers learners' questions.
  • Backend: Convex runs the actions that call Claude, Cartesia, and Elastic, plus the lesson cache and learner progress for the impact page.
  • Input and extension: Web Speech API for mic questions, and a Chrome MV3 extension for making lessons from any page.

How we keep it accurate

The first question anyone asks about AI health education is what happens when the AI is wrong. HerHealth has five layers:

  1. Trusted sources only. We teach from an allowlist: NIH, MedlinePlus, CDC, womenshealth.gov, ACOG, WHO, FDA, Mayo Clinic, Cleveland Clinic, and NHS. If you paste a link from any other site, we never teach from that page. We work out its topic and teach it from MedlinePlus instead, and the lesson tells you so.
  2. Grounded writing. Claude may only use facts stated in the retrieved passages. Each fact cites its source and quotes one sentence from it word for word.
  3. Quote check in code. Every quote must really appear in its source passage. Any fact or quiz item that fails is dropped.
  4. A second reviewer. A separate Claude call checks the right answer, the facts, the quiz items, the red flags, and the conditions to consider against the sources. Anything unsupported is removed. If the answer itself isn't supported, no lesson is made at all.
  5. Safety by design. The app never diagnoses. Urgent conditions show a call-911 banner, and every page says it is education, not medical advice.

Challenges we ran into

  • Trust versus coverage. Many health pages block automated requests, and some sites aren't trustworthy. We built a fallback: work out the topic from the link itself and teach it from MedlinePlus, with a relevance check so a loose search can't swap in another subject.
  • A teacher that feels alive on a phone. Browsers block audio that starts without a tap, the 3D model is heavy, and lip sync drifted ahead of the voice. We drove the mouth from the real audio signal, delayed the model load so text shows first, and added a tap-to-start fallback.
  • Respectful language. Early drafts leaned on stereotypes, like the picture of a typical ADHD kid. We rewrote the lesson rules so the contrast is always about symptoms, never about a type of person.

Accomplishments that we're proud of

  • A full lesson on any health topic in about 30 seconds, with every claim checked against its source.
  • A clear answer to "what if the AI is wrong": allowlisted sources, quotes verified in code, and a second reviewer.
  • One app that works for the general public and for medical students.

What we learned

Grounding is more than putting sources in the prompt. It takes checks after generation that can say no. We also learned how much a voice and a face change how people engage with health information.

What's next for HerHealth

  • Lessons from videos, using their transcripts.
  • More languages and voices.
  • Review of the built-in lessons by clinicians.
  • Class mode, so instructors can assign lessons and see results.

Built With

Share this project:

Updates

Submission history