Key features breakdown

What a job seeker can actually run in this repo, and which files back it.

• Resume analysis and structured extraction
• ATS scoring against a job description
• Tailored resume, enrichment, improvement, Unslop
• Cover letter, cold mail, LinkedIn drafts
• Interview practice and career coach
• PDF resume export
• Metered plans via Razorpay

This is not an employer screening suite. TalentSync-HR owns that. /dashboard/recruiter in the frontend is leftover chrome, not the HR product.

• Frontend: Next.js 16, Bun, Prisma for payments/session hydration
• Backend: FastAPI 3.13, LangChain, LangGraph, FastStream workers
• Jobs: Kafka lanes in backend/app/stream/topics.py
• Admin: optional operator console, Compose profile admin

graph TB
subgraph "Frontend Next.js"
FE_Seeker["Seeker dashboard"]
FE_ATS["ATS page"]
FE_Coach["Career coach"]
FE_BFF["app/api BFF"]
end
subgraph "Backend FastAPI"
BE_Resume["routes/resume_analysis.py"]
BE_ATS["routes/ats.py"]
BE_Jobs["stream + jobs table"]
end
subgraph "Workers"
W["APP_ROLE=worker"]
end
subgraph "Persistence"
DB["PostgreSQL"]
K["Kafka"]
end
FE_Seeker --> FE_BFF
FE_ATS --> FE_BFF
FE_Coach --> FE_BFF
FE_BFF --> BE_Resume
FE_BFF --> BE_ATS
BE_Resume --> BE_Jobs
BE_ATS --> BE_Jobs
BE_Jobs --> K
K --> W
W --> DB

• Resume analysis

  • PDF/text via PyMuPDF
  • Structured fields, then LLM notes
  • BFF 202, then POST /api/analysis/finalize after the job

• ATS

  • Resume plus JD text or URL
  • Score, missing keywords, suggestions
  • Assessment lane on Kafka

• Rewrite tools

  • Tailored resume, JD editor, AI resume editor
  • Enrichment / refine / regenerate
  • deslopify is queued Unslop, not a sync formatter

• Outreach

  • Cold mail, cover letter, LinkedIn post/page
  • Communication and social lanes

• Interview and coach

  • Session create, answer eval, summary on the interview lane
  • Token streaming stays on the API process
  • Career coach uses the primary LLM role; titles use the small role

• Billing

  • Razorpay subscriptions and credit packs
  • Backend feature_gate is the only meter. The BFF must not double-charge

sequenceDiagram
participant User as "Job Seeker"
participant UI as "Frontend"
participant API as "Next.js BFF"
participant BE as "FastAPI"
participant K as "Kafka"
participant W as "Worker"
participant DB as "PostgreSQL"
User->>UI : Upload resume or paste JD
UI->>API : POST feature route
API->>BE : fetchBackend
BE->>DB : job QUEUED
BE->>K : JobEnvelope
BE-->>API : 202 job_id
API-->>UI : wait on job
K->>W : consume
W->>DB : result
UI->>API : GET job
API-->>UI : payload

Accepts PDF and text-like uploads. ZIP bulk ingest is not the seeker happy path; treat multi-file ZIP as leftover from older hiring copy. Analysis is metered and async.

flowchart TD
Start(["Upload Resume"]) --> Parse["Parse with PyMuPDF"]
Parse --> Enqueue["Enqueue analysis job"]
Enqueue --> LLM["Worker LLM extract + notes"]
LLM --> Finalize["Client finalize persist"]
Finalize --> End(["Show results"])

Needs a resume and a JD (text or link). Failures are usually an empty JD or a job still RUNNING.

Cover letter and cold mail share the communication lane. LinkedIn uses the social lane. PDF generation is a frontend/LaTeX path (utils/latexGenerator.ts), not a Kafka LLM job.

Interview graphs live under backend/app/services/interview/. Sessions are Postgres (talentsync_backend.interview_session), not a process dict. Coach chat streams; title generation uses SMALL_LLM_*.

lib/plans.ts and app/services/metering.py define per-plan caps. Heavy features cost extra credits. 402 PAYWALL is the over-quota body.

graph LR
FE_UI["Frontend UI"] --> FE_API["Next.js BFF"]
FE_API --> BE_ROUTES["FastAPI Routes"]
BE_ROUTES --> JOBS["job submit"]
JOBS --> K["Kafka"]
K --> W["Workers"]
W --> SVC["Services"]
SVC --> DB["PostgreSQL"]

Long LLM calls do not block /api/v1/auth/me. Queue latency is created_at -> started_at on the job row. Consumer lag on ts.worker.resume is the alert that matters.

• Unsupported file: PDF or text. If the worker never starts, Kafka/worker is down.
• Empty resume text: the parser got no extractable text. Try another PDF.
• LLM 503: KAFKA_ENABLED false, or no worker.
• ATS: provide JD text or a reachable link.
• Paywall 402: check plan limits, not a model outage.