Cadence

One in four congestive heart failure patients is readmitted within 30 days of leaving the hospital. Most of those readmissions are preventable. The window right after discharge is when patients are most at risk, most confused by their new medication schedules, and least likely to call their doctor about a warning sign they're not sure is serious. We wanted to build something that could sit in that window with them.

The core idea came from thinking about who patients actually trust. Not an app. Not a notification. Their daughter. Their spouse. So we built Cadence around that: a voice clone of the caregiver that calls the patient each morning, asks about weight and breathing and meds, and quietly alerts the family when something looks off. The patient never has to open an app or navigate a screen.

We built the stack across three layers. The mobile app handles onboarding for caregivers, photographs discharge paperwork and runs it through Gemma 4 to extract a structured medication regimen, records the caregiver's voice, and then hands the phone to the patient. The backend handles daily log storage, a rule-based escalation engine, and ElevenLabs TTS to synthesize the caregiver's voice for each prompt. The caregiver dashboard gives a full 30-day view of weight trends, medication adherence, and any flagged days.

The hardest part was the document processing pipeline. Discharge paperwork is messy, inconsistent, and often photographed under bad lighting on a phone. We ran Apple Vision on-device to extract raw text, passed it through Zetic to strip personal info, then sent the anonymized output to Gemma to produce a structured medication list. Any failure in that chain meant the patient's regimen didn't get created, so every step had to be robust. Getting Gemma to return clean structured JSON without hallucinating doses or dropping medications took a lot of prompt iteration, and we hit a wall early on with the multimodal API returning empty responses when JSON mode and image inputs were combined, which cost us a few hours to diagnose. Getting the full chain from photo to anonymized text to parsed regimen to stored medications to voice-prompted check-in working end to end was the moment the project clicked into place.

Built With

• elevenlabs
• expo.io
• fastapi
• javascript
• mongodb
• python
• react
• typescript