KinBraid

One shared care reality, translated into the next right action for every person in the care circle.

Comment

Inspiration

Family caregiving rarely fails because people do not care. It fails because the relevant facts are scattered across medications, appointments, measurements, notes, and different people’s mental models. KinBraid turns that fragmented care reality into a calm, role-specific plan for what matters now.

What it does

Our OpenAI Build Week feature, Today's Care Brief, gives a family coordinator or caregiver a source-bounded daily briefing built from the patient data already in KinBraid. It summarizes the current situation, separates actions into Today and Next 48 hours, suggests focused questions for the care team, calls out when escalation may be appropriate, and shows the source changes behind every brief.

The result is deliberately not a chatbot and not a diagnosis. It is a structured coordination layer that helps a care circle understand the next right action without pretending to replace a clinician.

How we built it

The product uses React, TypeScript, Vite, Express, PostgreSQL, and Capacitor. The Care Brief is generated with GPT-5.6 through the OpenAI Responses API using strict Structured Outputs. Requests run on the Flex processing path with a longer timeout and conservative retry-ready behavior. The application validates model output against the same schema before rendering it.

Each request receives a bounded evidence packet containing only the relevant care-circle records. The UI returns a stable care-plan shape with a status label, summary, Today and Next 48 Hours actions, questions, escalation guidance, recent source changes, explicit uncertainty, and generation provenance.

How we used Codex

Codex was our primary implementation partner during Build Week. We used it to trace the existing care-state architecture, implement the source-bounded KinBraid Intelligence endpoint and strict schema, integrate GPT-5.6 Flex processing, repair a persistence bug uncovered during live testing, build the responsive role-specific UI, improve accessibility and safety language, write setup and architecture documentation, and validate the full flow with lint, production build, API smoke tests, and desktop/mobile browser checks.

Safety and trust

KinBraid never presents the brief as medical advice. It avoids diagnosis and treatment changes, retains explicit uncertainty, keeps source changes visible, and uses escalation cues rather than clinical conclusions. Generated briefs are not treated as durable source-of-truth records; the underlying care data remains authoritative.

Challenges

The hardest part was making the output useful without letting it become speculative. We solved that by constraining both the model’s input and its output: bounded evidence in, strict typed care-plan data out, followed by server-side validation and a UI that keeps provenance and caution visible.

Flex processing also required operational changes. We gave the Responses API a longer request timeout and made the service tier and reasoning effort configurable while surfacing the actual generation path in the product.

Accomplishments

A real GPT-5.6-generated care brief from synthetic care data in roughly 16 seconds on Flex processing Strict Structured Outputs with server-side validation Role-aware coordinator and caregiver experiences Source disclosure, uncertainty language, and escalation boundaries Responsive desktop and mobile UI with accessible loading and touch behavior A reproducible judge path and documented Build Week work boundary

What is new for Build Week

The repository and core care-circle product existed before the event. The Build Week work adds the model-backed, source-bounded KinBraid Intelligence pipeline, strict output contract, role-aware daily briefing UI, GPT-5.6 Flex configuration and provenance, and the associated safety, documentation, and verification work.

What's next

Next we will add evaluation fixtures for source fidelity and role usefulness, richer evidence linking from each action to its originating record, and carefully scoped additional roles while preserving the same non-diagnostic boundaries.

Built With

Share this project:

Updates