Autonomous Career Agent (ACA) -- Phases 1-3

A resume parsing, ATS tailoring, and application-tracking tool, built as a single Next.js app.

Current engineering evidence and known limitations: ACA audit. Project explanation and interview preparation: Recruiter guide. Account protection and recovery: Two-factor authentication.

Scope vs. the original spec

The full product spec (Autonomous_Career_Agent_Specification.pdf) describes a much larger system: a separate Python FastAPI backend, Postgres + pgvector, Redis/BullMQ job queues, Gemini via Vertex AI, and a Playwright automation cluster that headlessly submits applications while evading anti-bot detection on Greenhouse/Lever/Workday/Ashby.

This build implements the same user-facing value entirely inside the Next.js app, with two deliberate deviations:

1. Next.js APIs with Firebase accounts. Processing runs in Next.js Route Handlers on Vercel. Firebase Authentication protects access and Firestore stores private profiles and applications across devices. See FIREBASE_SETUP.md for local and deployed configuration.
2. No stealth/anti-bot-evasion automation. The spec's Phase 3 calls for headless browser workers that rotate IPs and spoof fingerprints to auto-submit applications unattended. That's built specifically to defeat a platform's bot defenses, which is very likely a violation of those platforms' terms of service -- so it isn't included. Instead, Phase 3 here is an application tracker plus a fill-assist bookmarklet that pre-fills common fields (name/email/phone) on a job page inside your own browser tab, when you click it. You still review and submit yourself.

Phases

• Phase 1 -- Profile: upload a PDF resume; it's parsed server-side (app/api/upload/route.ts) into name/email/phone/skills plus the full extracted text.
• Phase 2 -- Tailor: paste a job description; app/api/tailor/route.ts computes a 0-100 ATS-style match score using cosine similarity over term-frequency vectors (Similarity = (A·B) / (||A|| ||B||), per the spec) blended with explicit skill-keyword coverage, then returns matched/missing skills and rule-based tailoring suggestions that never invent skills or experience you don't have.
• Phase 3 -- Track: a status pipeline for applications (discovered → tailored → applied → interviewing → offer/rejected), plus the fill-assist bookmarklet described above.

Optional AI-enhanced suggestions

The tailoring engine's score is always the deterministic heuristic above. Its suggestion phrasing can optionally use Anthropic after the user explicitly opts in. Paid calls require a signed visitor session, shared Redis rate/budget guards, and server-side credentials. Without all prerequisites, the app uses local suggestions.

See DEPLOYMENT.md for configuration, request/resource limits, account spend limits, monitoring, privacy, and the tracked dev-dependency advisory. Do not enable paid AI using an API key alone.

Getting Started

npm ci
npm run dev

Copy .env.example to .env.local and fill the Firebase web configuration first. Open http://localhost:3000 for the landing page. Its links lead to login/registration; verified users open /app.

Known limitations

• PDF text extraction (pdf2json) reconstructs line layout from text coordinates, which works well for standard single/two-column resumes but can still struggle on heavily designed, multi-column, or scanned/image PDFs (the spec's own noted Phase 1 risk).
• The skill vocabulary in lib/skills.ts and the ATS scoring in lib/ats.ts are heuristics, not a trained model -- treat scores as directional, not authoritative.
• Workspace data requires an online connection and a verified Firebase account. Existing browser data can be imported explicitly from Backups and data. Original PDFs are not stored; extracted resume text is saved in the account.