Anh
Trinh

An AWS serverless pipeline that turns Lever ATS activity into client-ready candidate scorecards, with Claude as the assessment layer. Handled around 50 candidates a month and saved each recruiter one to two hours of manual data entry, while keeping a human in front of every client-facing document.

• Distributed, event-driven pipeline in TypeScript on the Serverless Framework — five Lambda functions behind API Gateway, SQS with a dead-letter queue, DynamoDB and EventBridge — ingesting Lever webhooks and mirroring every candidate into Google Drive and Google Sheets.
• Duplicate output files traced to Lever retrying against the 29-second API Gateway timeout, fixed by splitting the fast signature-verifying endpoint from a 120-second SQS worker that does the CV parsing, image downloads and Drive writes.
• A Claude assessment step drafts the scorecard against a purpose-built JSON schema; a render Lambda turns the approved draft into a branded A4 PDF with Handlebars and Puppeteer and posts it to Slack. Gemini matches interview transcripts back to the right candidate.
• Every write is idempotent under at-least-once delivery: DynamoDB provisioning locks, Google Drive appProperties as the file identity rather than the file name, and checksum change detection. EventBridge sweeps retry a failed transcript match or PDF render; the DLQ holds anything still failing for 14 days.
• The core stays ATS-agnostic behind an adapter contract, covered by unit, integration and live end-to-end suites against the real Lever, Google and Slack APIs, plus a data-completeness test that fails the build if a candidate field goes missing between artifacts.