CareRelay

A traceable handoff layer for family care

CareRelay turns fragmented post-discharge information into clear, source-grounded family action. It is not a medical chatbot: it reconciles what the family has received, keeps uncertainty visible, assigns the next operational step, and creates a handoff another caregiver can actually use.

Its core loop is:

evidence -> conflict or gap -> assigned action -> handoff

What a reviewer can test

The public prototype uses only fictional people and synthetic documents. Select Sources and load the recommended four-source packet:

• discharge-summary PDF
• medication-list photo
• family-chat TXT export
• cardiology appointment-email TXT export

Then select Save source changes and Analyze.

PDF and TXT content is read locally into editable source context. JPG and PNG files are attached as server-side image evidence for GPT-5.6. The current packet becomes source-linked facts, unresolved conflicts, action tasks, gaps, and an updated handoff.

A reviewer can assign or complete a task, open View evidence to see the exact highlighted sources behind an action, and select Handoff -> Copy handoff to copy the generated brief to the real clipboard. If a browser blocks clipboard access, the modal stays open with the complete readonly handoff and a working Select all control for manual copy. The source packet, current analysis, and task state persist locally in the browser and are clearly marked as a fictional local demo workspace, not unavailable multi-user synchronization.

Why this is different

Existing tools often handle reminders, documents, or chat separately. CareRelay is the coordination layer between them: every important item should have a source, a responsible person, and a next action. It makes conflicts actionable without pretending to resolve a medical decision.

High-risk choices remain unresolved until a qualified human confirms them. The product never diagnoses, chooses a medication plan, or turns ambiguity into false certainty.

GPT-5.6 and Codex

The deployed server-side GPT-5.6 endpoint returns strict structured JSON for facts, conflicts, tasks, and gaps. The React interface maps that live result directly into evidence, actions, and handoff content. The implementation accepts bounded image evidence, caps inputs, uses low reasoning effort for this targeted extraction task, rate-limits live calls, times out upstream requests after 55 seconds, and caches identical fictional packets to avoid repeat demo costs.

Codex was used for product scoping, React and TypeScript implementation, the structured analysis contract, responsive workflow, cost controls, safety constraints, synthetic evidence design, local persistence, document-input testing, clipboard success/failure verification, and browser verification.

Safety and privacy

The demo uses fictional data only. CareRelay does not diagnose, replace a clinician, alter medication instructions, or infer a medical decision from conflicting information. A production version would require consent, encryption, access controls, audit logs, clinical governance, and authenticated multi-person collaboration.

Built With

• gpt-5.6
• openai-responses-api
• react
• typescript
• vercel
• vite