Bedside
A voice-first handoff app for home hospice volunteers and family caregivers. Bedside turns spoken shift notes into structured handoffs, helps the next person understand what changed, and flags concerns that may need a nurse's attention.
2nd place overall
Making handoffs easier for the people providing care at home.
Home hospice care involves nurses, family members, and volunteers sharing responsibility across different shifts. Important context can live in handwritten notes: medication updates, changes in behavior, what helped, and what the next person should know. We wanted to explore what the handoff could look like if someone at the bedside could simply talk through what happened. Bedside turns that voice note into a structured shift summary and uses the same information to help the next caregiver understand what changed.
Medicare-certified hospices are required to maintain volunteer activity equal to at least 5% of the patient-care hours provided by paid staff. Volunteers are not an edge case in hospice care. They are a required part of how hospice programs operate. Source: CMS.
Sydney focused on product design and the React frontend. I worked on the backend, data model, infrastructure, and three-agent architecture connecting everything together.
Handoff summarizer
Turns a spoken shift note into structured information including medications mentioned, mood, what helped, urgency, and confidence. If a volunteer's handoff contains a potential red flag, Bedside alerts the primary caregiver and asks whether they want to notify the nurse or care team. The system surfaces the concern, but the caregiver makes the final decision.
Care-plan Q&A
Uses the household's care plan to answer questions and show where the information came from. If the answer is not available, it directs the user to the nurse instead of generating one.
Pattern analyst
Reviews recent shift logs together to identify patterns that may be difficult to notice from a single handoff. In our synthetic data, it surfaced changes such as sundowning beginning earlier and signs of caregiver strain across multiple shifts.
Stored agent responses as JSONB so the response structure could evolve without requiring schema changes during the hackathon.
Kept DigitalOcean agent credentials server-side so no keys were exposed to the browser.
Parsed agent responses defensively and returned a safe fallback when an output was malformed instead of crashing the interface.
Worked through issues including escalation flags being downgraded, tool calls being suppressed by the agent prompt, and a session logout bug after switching roles.
Sydney focused on product design and the React frontend while I worked on the backend, agent system, and integration.
All three agents ran live using synthetic patient data. No real patient information was used.
15 hours from idea to presentation.
Planning, building, and presenting. The hour splits inside the 15 are approximate.
Working in parallel: while the inference provider was returning 403 errors, Sydney kept building the frontend against mocked responses, so both tracks stayed productive until live agent responses returned.
What we ran into, and what I'd change.
Inference calls returned 403 errors
Our inference calls began returning 403 errors even though authentication was working correctly. After debugging the account and endpoints, we determined the issue was outside our application.
An agent followed formatting instructions too literally
One agent followed my JSON formatting instruction so literally that it suppressed the tool call used to surface an escalation prompt to the caregiver.
The frontend and backend used different data
The frontend and backend were initially using different patient data. Both worked independently, but the mismatch became obvious when we connected them.
A handoff designed around the shift itself.
Simple entry
Scan the household QR code, choose a profile, and enter a PIN without creating an account or downloading an app.
Voice handoff
Describe what happened during the shift instead of filling out a structured report.
Shift context
See what changed, review the care plan, and understand routines before starting the next shift.
Role-based information
Family members, volunteers, and nurses see information based on what they need for their role.
Bedside was one of the first projects where I spent most of my time designing what happened behind the interface. The agents were only one part of the system. I had to think about what data they could access, how their responses were structured, when they should escalate, what happened when they failed, and how the frontend could keep moving while the backend was unavailable. It changed how I think about AI products. The interface is only one layer of the experience. The behavior underneath it needs to be designed too.
Two designers, 15 hours from idea to prototype.
Michelle Bui worked on the backend and agent architecture. Sydney focused on product design and the React frontend.
