TalentSync

AI resume builder and ATS optimizer for job seekers. Tailored resumes, outreach, and interview prep from one profile.

Source

Introduction and problem statement

TalentSync is for people applying to jobs, not for HR teams screening a pile. ATS filters still drop qualified resumes over formatting and missing keywords. Rewriting a resume per posting by hand is slow and you get no signal.

Live: talentsync.tashif.codes

TalentSync-HR is a different product. Employer dashboards and bulk recruiter ingest belong there.

Problem statement#

Job seekers write into a black box. Parsers want a shape humans do not write. Feedback is generic. Each application wants a slightly different document.

The old README quotes 49 applications per job and 93% ATS adoption as recruiter pain. This codebase spends its routes on the seeker side of that mess: parse, score, rewrite, draft outreach, practice interviews.

Pain on the seeker side#

  • ATS rejection for formatting or keyword layout, not skill.
  • Vague “we went another direction” with nothing to change.
  • Manual tailoring per posting.
  • No score that says why this resume missed that JD.

What this app does#

  • Resume analysis with structured extraction (PyMuPDF, LLM).
  • ATS evaluation against a job description.
  • Tailored resume, JD editor, AI editor, enrichment, Unslop.
  • Cover letters, cold mail, LinkedIn drafts.
  • Interview sessions and a career coach.
  • Metered plans (Razorpay). Free/paid caps live in frontend/lib/plans.ts and backend/app/services/metering.py. Core tools are not “unlimited free” in the current metering code.

What it does not do#

  • Rank a company’s inbound applicants.
  • Replace an ATS for employers.
  • Ship recruiter bulk ZIP screening as the headline feature.

Technical foundation#

LangChain / LangGraph for generation. FastAPI enqueues metered AI to Kafka. Workers write results to talentsync_backend.job. Next.js is the UI and BFF. Google OAuth is the default login.

The scikit-learn pickle under analysis/ is a leftover classifier. Production analysis does not go through that notebook.

How a request lands#

  1. Browser hits a Next.js route.
  2. BFF authenticates via backend JWT cookies and forwards to FastAPI.
  3. feature_gate charges, then submit_job inserts a row and publishes a thin envelope.
  4. Worker runs the existing service function.
  5. Client waits on job_id (poll or SSE).

Token streams skip that queue on purpose.