Solace is an AI patient intake and triage app for emergency departments. Here is the full stack and why each piece is where it is. The whole thing runs on AWS managed services with the human-facing apps on Vercel.
Patient phone / Clinician (Vercel SPA)
-> CloudFront + WAFv2
-> API Gateway (HTTP)
-> Lambda (FastAPI, arm64 container)
-> DynamoDB, S3, Bedrock, Transcribe, Polly
Cross-cutting: KMS CMK, Secrets Manager, CloudTrail, CloudWatch, EventBridge, SNS.
A patient checks in by scanning a QR code in the waiting room, with no app and no account. The patient app and the clinician terminal are static, edge-cached shells on Vercel that load fast on a weak waiting-room connection, then call the AWS backend. Branch-based atomic deploys mean I can ship the marketing site, the patient app, and the clinician app independently.
One FastAPI service (via Mangum) packaged as an ARM64 container image on AWS Lambda, 2048 MB, 60s timeout, Python 3.12 on AL2023, behind an HTTP API Gateway. The trained ML artifacts are baked into the image so cold starts do not pay a model-download tax. The registry is ECR.
Amazon DynamoDB is the primary datastore: about 30 tables, on demand, CMK-encrypted, with TTL on transient state (idempotency keys, session nonces, hot audit records). The live queue is a GSI on hospital_id + created_at
. Media lives in S3 with presigned delivery and public access blocked.
Because it is healthcare:
alias/solace
) encrypts every DynamoDB table, the media bucket, the CloudTrail bucket, and Secrets Manager.The point of the architecture is that a managed-service stack lets a small team ship something HIPAA-grade that still scales, rather than choosing between fast and real.
I created this content for the purposes of entering the H0: Hack the Zero Stack hackathon. #H0Hackathon