Captured from the live deployment, not a mockup
01 · The problem
Built the AI screening path of a hiring platform: a resume uploads to S3 on a presigned POST, an SQS-fed Lambda has Bedrock Nova Lite score it through a forced tool call that treats the resume as untrusted input, and a transactional outbox emails the candidate once.
02 · How it works
- 01
Measured the model instead of trusting it: on 130 invented resumes strong ones averaged 73 and weak ones 25, strong ahead in all 1,612 pairs, and a resume telling the model to output 100 scored 20. That shows the score tracks evidence, not agreement with recruiters.
- 02
NestJS and PostgreSQL API with a React console, defined in AWS CDK, deployed over GitHub Actions OIDC: 609 tests, 97% API statement coverage.
- 03
Built a hiring platform as a NestJS and TypeScript API on PostgreSQL with a React console: an illegal pipeline jump returns 409, a stale edit is refused by optimistic locking, and long lists page by keyset instead of offset.
- 04
Resumes upload to S3 under a presigned POST that enforces size, type and key; an SQS-fed Lambda scores each with Bedrock Nova Lite, and a transactional outbox sends each email once. On 130 invented resumes strong ones averaged 73 and weak 25, strong ahead in all 1,612 pairs; a resume telling the model to output 100 scored 20.
03 · What it cost, and what it returned
Defined the stack in AWS CDK (EC2, RDS, S3, CloudFront, SQS, Lambda, SES, 12 alarms) and deploy it from GitHub Actions over OIDC with a 28-check smoke test per release; 609 tests, 97% statement coverage on the API.