Symora

study team/medical personnel side of the dashboard where they access patient symptoms
user facing app, home screen
sign in page for the user
drop down for user to select their preferred language
main page for user to submit their symptoms
profile, user can either sign out or change their preferred language
user symptom screen, they can fill out any gaps that their chat/voice message missed
post-submission screen
notification center for the user to receive news based on their symptom, to either follow-up with a doctor or if they're good to go

Inspiration

Clinical trials depend on participants accurately reporting side effects. But someone who doesn't speak English well may skip a report, downplay a symptom, or describe it in a way staff can't capture: "my stomach is burning and my head is spinning." Every missed or garbled report is a hole in the safety data, and participants who feel unheard are more likely to drop out.

We asked a simple question: what if a participant could just talk, in their own language, and the study team received a clear, verified English record?

We don't try to get people into trials. We make sure the people who are already in can be heard.

What it does

Symora is a voice reporter for clinical trial participants that translates the language they are most comfortable speaking into a structured English report for clinical trial use.

The participant speaks in their own language (Spanish and Vietnamese in our demo).
A voice agent asks short follow-up questions: when did it start, how bad is it, is it still happening?
Gemini converts the conversation into a structured English report: symptom, onset, intensity, and ongoing status.
The participant hears a read-back and confirms before anything is submitted.
The confirmed report is saved to Postgres for the study team.

Symora is not a translation app. A translator converts words; Symora produces a trial record. Unknown details stay null, and we never guess a date, a severity, or a medication.

How we built it

Gemini (structured JSON output) extracts and translates symptoms from the transcript, with a fallback model for when the primary model is busy.
ElevenLabs powers the voice conversation, the entire participant interface.
Tiger Data (PostgreSQL) stores participants, symptoms, and reports as timestamped records.
Next.js serves the app and the API, with bearer-token authentication on every endpoint.

Challenges we ran into

Making a language model trustworthy. Models can add details nobody said, so we built verification around the model instead of relying on the prompt alone.
Real behavior vs. tests. Mocked tests passed while the live API still had bugs, so we learned to verify with real requests.
Reliable multilingual support. Each language needed testing of its own, so we limited our demo to the languages we could verify.
Teamwork across a fast-moving codebase. Several people built pieces at once, so we used branches and pull requests to keep everything working together.

What we learned

Safety-critical AI needs verification layers around the model, not just a good prompt.
Test with real requests, not only mocks.
Being honest about what doesn't work yet builds more trust than hiding it.

What's next

A coordinator dashboard that highlights urgent reports, with a human reviewing every report
Support for more languages, including Mandarin
Routing each report to the coordinator assigned to the participant's study site
Real authentication and role-based access, plus HIPAA and IRB review before any real deployment

Safety and privacy

Symora is a prototype built with fictional data only. It gives no medical advice and never diagnoses. It stores a participant ID and nothing else: no names, and no immigration status. The original transcript is always preserved, and the participant confirms every report.