Bedrock Triage Agent

A lightweight AI-powered triage tool that classifies support issues by category, severity, and summary using Amazon Bedrock (Claude 3 Haiku).

Comment

Inspiration

Triage is where teams lose the most time in incident & bug flow. I wanted a tiny, reliable helper that reads a free-text report and instantly proposes Category, Severity, and a crisp Summary — with or without cloud access.

What it does

Bedrock Triage Agent takes a plain-English description (e.g., “Checkout page fails with NullPointerException…”) and returns:

  • Category: BUG / FEATURE / QUESTION
  • Severity: CRITICAL / HIGH / MEDIUM / LOW
  • Summary: One-line, actionable title

It ships with:

  • A lightweight web UI (Thymeleaf) at /
  • A simple JSON API at POST /api/triage
  • Two profiles:
    • mock → fully offline, deterministic results (no AWS needed)
    • bedrock → real call to Amazon Bedrock (Claude 3 Haiku) via AWS SDK v2 + SSO

How I built it

  • Spring Boot 3 (Web, Actuator, Thymeleaf)
  • AWS SDK v2 (bedrockruntime, sso, ssooidc)
  • Clean service split:
    • ApiController → REST endpoints & UI routes
    • TriageService → calls Bedrock (or MockTriageService for offline)
    • BedrockConfig → builds the Bedrock client based on profile/region
  • Prompt engineering to map model output into structured fields.

Why it’s functional for judging

  • One-command local run (no AWS required):
    mvn -q spring-boot:run -Dspring-boot.run.profiles=mock
  • Real Bedrock run (optional):
    1) aws configure sso --profile bedrock-sso
    2) mvn -q spring-boot:run -Dspring-boot.run.profiles=bedrock

Challenges

  • Getting SSO + Bedrock working cleanly with the Java SDK.
  • Designing a prompt that is short, consistent, and easily mappable to enums.
  • Providing a deterministic offline mode so anyone can try it instantly.

Accomplishments

  • End-to-end triage working from UI and API.
  • Runs with or without AWS access.
  • Clear README + quickstart + architecture diagram.

What I learned

  • Practical Bedrock integration patterns in Java.
  • Small prompt tweaks dramatically change consistency.
  • Keeping the API contract tight makes the UI trivial.

What’s next

  • Multi-label categorization & confidence scores.
  • Slack/Jira integration for one-click ticket creation.
  • Docker image + ECS/Fargate deploy.

Built With

Share this project:

Updates