Skip to content

aws_deployment.md

Latest commit

 

History

History
62 lines (45 loc) · 1.91 KB

aws_deployment.md

File metadata and controls

62 lines (45 loc) · 1.91 KB
Raw

AWS Deployment

The deployment target is AWS App Runner running two Dockerized services:

  • FastAPI inference API
  • Next.js healthcare risk console

This keeps the production path small enough to maintain while still showing a real cloud stack: Docker, ECR, managed HTTPS, health checks, Terraform, and token-protected APIs.

Deployment Flow

  1. Train the model on Kaggle.
  2. Download risk_model.joblib into artifacts/.
  3. Create the ECR repositories with Terraform.
  4. Build the API and web Docker images.
  5. Push both images to ECR.
  6. Deploy both App Runner services with Terraform from infra/aws/apprunner.

Runtime Contract

The container expects:

  • MODEL_ARTIFACT_PATH=/app/artifacts/risk_model.joblib
  • REQUIRE_API_TOKEN=true
  • API_TOKEN=<strong-token>
  • OPENAI_API_KEY=<optional-key>

The Next.js container expects:

  • API_BASE_URL=https://<api-service-url>
  • API_TOKEN=<same-token-used-by-api>

If the model artifact is not present, the service starts with the local demo CSV model. For the portfolio deployment, include the Kaggle artifact in the image before pushing to ECR.

API Checks

curl -s https://<service-url>/health
curl -s https://<service-url>/v1/predict/summary \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -d '{"age":52,"height_cm":170,"weight_kg":88,"bp":142,"glucose":160}'

Web Check

Open the App Runner web service URL and run a patient assessment from the browser. The browser calls the Next.js /api/predict route, and that server-side route forwards to the protected FastAPI /v1/predict endpoint.

Portfolio Talking Points

  • reproducible Kaggle training
  • persisted model artifact with metrics and model card
  • FastAPI inference service
  • Next.js production risk console
  • Streamlit decision dashboard as a local fallback
  • retrieval, memory, safety checks, and explainability
  • Dockerized deployment to AWS App Runner with Terraform