Pulse

Challenge

Care coordination often lives in inboxes and disconnected tools. Important follow-ups depend on someone remembering the thread. The product had to reduce reconstruction work without pretending the AI is the clinician.

Context

The concept is built for operations teams who handle high volumes of similar, time-sensitive tasks. Accuracy, auditability, and a clear human handoff mattered more than novelty.

Strategy

We designed Pulse around a queue of work, not a chatbot. AI drafts, retrieves, and suggests. People confirm, override, and own the outcome. Retrieval is scoped to authorized records, and every suggestion is inspectable.

Design approach

The interface is calm and dense where it needs to be. Status, next action, and risk are visible before decorative wellness imagery. Mobile is treated as a first-class operations surface, not a reduced dashboard.

Architecture

A React application talks to a FastAPI service. Retrieval-augmented generation sits behind explicit tool contracts. Source documents, prompts, and outputs are logged so a suggestion can be traced.

Implementation

We started with one workflow: follow-up after an encounter. Once that loop was reliable (retrieve, draft, confirm, record), adjacent workflows could reuse the same pattern.

Solution

Pulse gives coordinators a single operational canvas. Automation handles the repetitive assembly of context. People keep judgment, and the system keeps a record of why an action was taken.

Results

• A production-shaped pattern for AI in operations, not a chat demo.
• Human confirmation remains in the loop for any action that affects a person.
• A foundation for additional workflows without a new product each time.

Learnings

• AI is most useful when it prepares work, not when it hides the work.
• Logging the suggestion is as important as generating it.