Job Tracker

A job search is mostly repetitive work: reading alert emails, checking careers boards, judging whether a posting fits, researching the company, tailoring a resume and a cover letter, and remembering to follow up. Job Tracker started as the place where every application lives — where it came from, which documents went out, the answers given, the status and the next step — and grew into a system that hands that repetitive work to AI while I keep the decisions.

Discovery and fit scoring. Every hour, a scheduled function polls the job feeds whose terms allow it and the careers boards of the companies I follow (Greenhouse, Lever, Ashby, Workable, Recruitee, Teamtailor). Postings are deduplicated and filtered by location, then Claude on Amazon Bedrock scores each one against my verified profile: a fit score out of 100, the skills that match, the gaps, and whether those gaps look learnable. Pasting a posting into the app works too, and AI fills in the company, title, salary, deadline and requirements.

AI agents for the tasks around each lead. The app runs seven agents on Anthropic's Claude Managed Agents, each with one job:

• an alerts reader goes through the daily job-alert emails from boards that ban automation (LinkedIn, Indeed, Glassdoor…) over a scoped, read-only Gmail connection and turns them into scored leads;
• a researcher searches the public web to check a company, its posting and the people named in it, and returns a verdict with its sources — a scam check before I spend any time;
• a writer prepares a tailored resume, a cover letter and screening answers from my confirmed facts only, tracing every claim to a fact;
• a follow-up reader matches recruiter replies to applications and proposes the new status, note or next step;
• a learning planner turns the skill gaps that keep appearing into dated study plans;
• plus a general assistant and a profile drafter for a job platform's API.

Humans approve, agents propose. Nothing an agent does confirms, submits, sends or deletes. Every change arrives as a proposal I accept or reject, risky commands become approvals, and each session runs under a dollar budget that Anthropic enforces as a hard cap. The app runs each agent's custom tools itself, against my data and scoped to my account, and only data I have confirmed reaches an agent. Phone, address and salary are shared only when I allow it for that session.

The platform side is AWS Amplify Gen 2 end to end: Cognito for the single owner account, AppSync and DynamoDB for owner-only data, S3 for the document library, Lambda functions for discovery and for the agent webhook, and EventBridge schedules for the hourly poll and the daily email reads. It is built spec-first, one GitHub issue at a time, and each feature is verified with Playwright end-to-end tests before it ships.