Project overview

TalentSync is a job-seeker product: resume analysis, ATS scoring, tailored rewrites, outreach drafts, interview practice, and a career coach. Live at talentsync.tashif.codes

It is not TalentSync-HR. That repo is recruiter screening. Do not document bulk employer shortlists or an HR talent pool as this app.

The upstream README still talks about dual-sided hiring and clones harleenkaur28/AI-Resume-Parser. Ignore that. This tree is tashifkhan/TalentSync.

graph TB
subgraph "Frontend Layer"
FE_Next[Next.js Frontend]
FE_UI[React Components]
FE_API[BFF app/api]
end
subgraph "Backend Layer"
BE_FastAPI[FastAPI API role]
BE_Worker[FastStream worker]
BE_AI[LangChain / LangGraph]
end
subgraph "Data Layer"
DB_PostgreSQL[PostgreSQL]
DB_Prisma[Prisma public]
DB_Jobs[Alembic talentsync_backend]
K[Kafka]
end
subgraph "Infrastructure"
INF_Docker[Docker Compose]
end
FE_Next --> BE_FastAPI
FE_UI --> FE_Next
FE_API --> BE_FastAPI
BE_FastAPI --> K
K --> BE_Worker
BE_Worker --> BE_AI
BE_Worker --> DB_Jobs
BE_FastAPI --> DB_PostgreSQL
DB_Prisma --> DB_PostgreSQL
INF_Docker --> BE_FastAPI
INF_Docker --> FE_Next
INF_Docker --> DB_PostgreSQL
INF_Docker --> K

Compose local services: db, kafka, backend (APP_ROLE=api), worker (APP_ROLE=worker), frontend. Admin and observability are Compose profiles, not the seeker app.

AI pipelines#

LangChain / LangGraph behind FastAPI routes. Workers run the slow ones. Streaming chat and interview SSE stay on the API process.

Seeker tools that enqueue:

• Resume analysis, improvement, enrichment, tailored resume, Unslop (deslopify)
• ATS evaluation and tips
• Cover letter, cold mail, LinkedIn
• Interview session create / answer eval / summary
• Hiring-assistant answers (practice copy, not an employer inbox)

Resume analysis#

PDF/text through PyMuPDF, then structured extraction, then LLM feedback. The BFF returns 202; the client waits on the job, then finalize persists the row.

ATS optimization#

Keyword overlap, compatibility score, and concrete edits against a job description. Same Kafka assessment lane as tips.

Career coach#

Primary LLM role for long answers. Small LLM role (SMALL_LLM_*, default gemini-3.1-flash-lite) for conversation titles only.

sequenceDiagram
participant Client as "Browser"
participant Frontend as "Next.js BFF"
participant Backend as "FastAPI api"
participant Kafka as "Kafka"
participant Worker as "FastStream worker"
participant Database as "PostgreSQL"
Client->>Frontend : Use a seeker tool
Frontend->>Backend : POST /api/v1/...
Backend->>Database : INSERT job QUEUED
Backend->>Kafka : JobEnvelope
Backend-->>Frontend : 202 job_id
Frontend->>Backend : poll or SSE
Kafka->>Worker : consume lane
Worker->>Database : result SUCCEEDED
Backend-->>Frontend : job result
Frontend-->>Client : Render

Design that actually shows up in the repo:

• One backend image, APP_ROLE selects api / worker / migrate
• Prisma owns public; Alembic owns talentsync_backend
• BYOK keys are encrypted on the job row, never in Kafka
• Google OAuth on FastAPI. NextAuth is gone

flowchart TD
Start([Resume + JD]) --> Parse["Parse resume text"]
Parse --> Extract["Structured skills and experience"]
Extract --> Analyze["Enqueue ATS job"]
Analyze --> Score["Worker: score + keyword overlap"]
Score --> Optimize["Suggestions"]
Optimize --> Output["Return via job_id"]

analysis/ still has a scikit-learn pickle (best_model.pkl, TF-IDF) from the older classifier. Live product paths go through LangChain services, not that notebook, unless you explicitly run analysis/app.py.

classDiagram
class CareerCoach {
+stream(messages)
+title(small LLM role)
}
class ResumeAnalyzer {
+extract(text)
+analyze(resume)
}
class ATSService {
+evaluate(resume, jd)
}
CareerCoach --> ResumeAnalyzer : "uses profile"
ATSService --> ResumeAnalyzer : "uses extracted text"

No employer talent marketplace. No ZIP-of-hundreds recruiter ingest as the main story. /dashboard/recruiter in this frontend is leftover navigation, not TalentSync-HR.

graph TB
subgraph "Frontend"
FE_Next[Next.js 16.1.6]
FE_React[React 18.2.0]
FE_Query[TanStack React Query]
FE_Prisma[Prisma 6.19]
end
subgraph "Backend"
BE_FastAPI[FastAPI]
BE_LangChain[LangChain]
BE_LangGraph[LangGraph]
BE_FS[FastStream Kafka]
BE_Motor[Motor]
BE_PyMuPDF[PyMuPDF]
end
subgraph "Infra"
DB_PostgreSQL[PostgreSQL 16]
KAFKA[Kafka 3.9.1 KRaft]
end
FE_Next --> BE_FastAPI
BE_FastAPI --> BE_FS
BE_FS --> KAFKA
BE_LangChain --> BE_LangGraph
BE_FastAPI --> DB_PostgreSQL
FE_Prisma --> DB_PostgreSQL

• Frontend: Next.js, React Query, Radix, Prisma, Razorpay
• Backend: FastAPI, LangChain, LangGraph, Motor, PyMuPDF, FastStream
• Auth: backend Google OAuth, httpOnly JWTs, Next rewrite /api/v1/:path*

API stays thin: validate, meter, enqueue. Workers scale by lane in prod (worker_resume, worker_assessment, worker_comms). Enrichments can run for minutes; Kafka max.poll.interval is 30 minutes.

ATS scores look wrong. Fix keywords and format, re-run. The job must SUCCEEDED first.

Analysis 503. Kafka disabled or no worker. Set KAFKA_ENABLED=true and start APP_ROLE=worker.

Login fails. Google redirect URI and JWT_SECRET / BACKEND_JWT_SECRET must match. This is not a NextAuth misconfig.

Compose name clash. Prod file uses talentsync_* containers so TalentSync-HR (talentsync_hr_*) can share a Docker host.