Building Generative AI Applications in 4 Hours
I showed application builders how our team made dtax, a prototype that turns a spoken patient account into a summary for a clinician, during a four hour hackathon.
Amazon · Singapore · 25 March 2026
Four Hours to Make the Core Experience Work
I built dtax with Anni, Darryl, and Zhi Kai at Amazon, then shared the work at the Amazon Web Services (AWS) User Group. The prototype combines a voice interview with the selected patient's existing health information and prepares a summary for the doctor. A printable patient slip and a queue view connect the intake to the consultation.
The talk explains how we chose a small core experience, divided the work, and got it running online within the time available. This was a hackathon prototype, not a clinically validated service or a system ready for real patient data. Clinical validation and proper clinician access controls were outside the demonstrated build.
From Voice Intake to a Clinician Handoff
Voice and Patient Context Become a Ticket
The voice interview and selected patient's conditions, allergies, and medications fed a JSON ticket. Its fields included a summary, symptoms, assessment, recommended actions, clinician notes, and a P1 to P4 acuity category. These were prototype outputs prepared for a clinician, not validated clinical judgments.
Records and Queue State
Separate patient, conversation, ticket, and queue records connected the intake to waiting, in consultation, and completed states. The printable patient slip and queue view provided a handoff beyond the voice conversation. Keeping those records distinct separated the patient's existing information from the current interview and its generated summary.
Hosted Connections Within Four Hours
The React and TypeScript client used Amplify Data with DynamoDB and separate Bedrock paths for voice and ticket generation. I compared Amplify with Lightsail, a monolith with microservices, and templates with custom code against the four hour limit. Deploying early allowed the hosted model and data connections to be tested before the final minutes.
Cognito and Lambda appeared in the architecture discussion, but the published March snapshot used public API key authorization, not clinician access controls.