Inspiration

What it does

The app is a telehealth web platform for postpartum care, with three steps:

  1. A short questionnaire covering mood, sleep, anxiety, and support at home, including the ten-item Edinburgh Postnatal Depression Scale (EPDS).
  2. A dashboard that shows where a person's answers point, in plain language, with the evidence and sources behind each area.
  3. Matched support: therapists, support groups, doctors with bookable appointments, and hotlines, chosen to fit the person's dashboard.

Everything is framed as a screening summary, not a diagnosis

How we built it

The backend is a sequential pipeline of four AI agents. Each does one job and passes typed output to the next:

Intake analyzer: turns questionnaire answers into a structured risk profile. Research agent: pulls evidence from clinical sources for those risk areas. Comparison agent: compares the profile against the evidence and builds the dashboard data. Resource agent: matches the flagged areas to real providers and open appointment slots.

We used Python with FastAPI, Pydantic AI for the agents, and Claude Sonnet 5 as the model. One shared PipelineState object carries data between agents, and every agent's output is validated against a schema before the next agent sees it.

The EPDS gives each item a score and clinicians commonly treat totals around 10 to 13 or higher as a reason to follow up.

Challenges we ran into

  1. Safety comes before everything. Item 10 of the EPDS asks about thoughts of self-harm, and a model shouldn't be the only thing standing between a person in crisis and help.
  2. The model can raise the flag but can never clear it. If it's set, the pipeline skips the dashboard and shows a curated list of crisis resources right away.
  3. Keeping the model honest. The comparison agent may only use the evidence it was given, must cite it, and is instructed to say "your answers suggest…" instead of "you have…". Schema validation catches malformed output and retries.
  4. Real resources, not invented ones. The resource agent can only return providers its tools actually found, and hotline numbers come from a verifiable list.
  5. Privacy. Only the zip code, not the full set of answers, goes to the resource agent.

What's next for HerHealth

  1. Review of the dashboard language and thresholds by clinicians
  2. HIPAA-compliant hosting before any real user data
  3. Follow-up check-ins over the postpartum period, not a one-time screen
  4. More languages and more local resources
  5. Creating user profiles that can track long term guidance

Built With

  • claudeapi
  • fastapi
  • pydantic
  • python
  • streamlit
Share this project:

Updates

Submission history