PRODUCT ARCHITECTURE
Clara Care.
AI-powered healthcare application platform
See the system in motion.
Architecture walkthrough · 00:10
The application, backend authority, ephemeral AI access and event stream form a documented live interaction path.
Illustrative sequence based on technical documentation; it is not patient data or product footage.
Open video ↗A closer look at Clara Care.
01 ORIGINAL VISUAL
The public Clara Care site shows its recording and transcription widget beside a clinical form. The visible sample fields belong to the published demonstration; this is team-built product work.
React interface for live consultation workflows.
Live clinical sessions cross four trust boundaries.
A clinical session touches application state, user identity, a live AI conversation and external services at the same time. If the boundaries between them are vague, sensitive data ends up where it shouldn't.
The backend owns identity. The browser borrows access.
I lead the solution architecture within the team: API and service boundaries, AWS integration, persistence, session design and AI data flows. The consultation widget asks the backend for a short-lived token and uses it to connect to Gemini Live and to an event stream that sends analysis updates back. I also built the product surfaces that bring Clara into the clinic: a widget embedded in the clinic's own EMR that listens to the consultation and fills the existing form, a Chrome extension with a side panel, the clinician platform, and a native SwiftUI prototype for iPhone. Live session state is synchronized through S2 streams. Underneath runs a real-time speaker identification service: voice activity detection, ECAPA-TDNN voice embeddings and online clustering tell the clinical record who is speaking, while voiceprints stay encrypted and are never written to logs.
Credentials that expire by design.
Identity, short-lived credentials and real-time streams each need one clear owner across the browser, the backend and the external AI service. The backend keeps identity and issues access that expires, so the browser only ever holds temporary credentials.
Implemented flows, team-owned results.
The widget flows and platform components described here are implemented, as the platform's architecture documentation shows. Clara Care is a team product: its results belong to the team, and its operational data stays private.
Clara's platform architecture documentation. Code and internal links stay private. This page summarizes what those documents show.