Job Search Agent (Demo)
This is a sanitized, portfolio/demo replica of a private personal project.
The name shown throughout is real (this is my own project), but every contact
detail, employer, and work history entry is a fictional placeholder, no real
personal data is in this repository. The private original runs against a real
Gmail inbox and real application data, which is exactly why that one isn't
the thing being shown publicly.
An end-to-end, zero-budget job-search tool: ingests real postings (Indeed/Glassdoor/ZipRecruiter/Google via JobSpy, LinkedIn job-alert emails, and opt-in Irish job boards), scores them against a skill profile, tailors a CV per posting, tracks every application, and syncs status from Gmail replies, all on free tiers.
What this demonstrates
- Full-stack ownership: FastAPI + SQLAlchemy backend, React + Vite + TypeScript frontend, a multi-connector ingestion pipeline, and CI for all three, deployed end to end (Render + Vercel + Neon Postgres).
- Applied RAG: pgvector-backed semantic skill matching (
backend/app/services/rag.py) blended with exact keyword matching, used to decide which of a candidate's real skills are relevant to a given job description, no LLM call needed for that half of resume tailoring.
- LLM integration done defensively: a multi-provider abstraction (Groq primary, Gemini fallback, both genuine free tiers) where every call site degrades to deterministic rule-based logic if a provider is unconfigured or a call fails, the app never breaks because a third-party API is down.
- A hard-enforced no-fabrication guardrail: CV tailoring can only reorder or select from a candidate's own stored profile data, never invent a skill or metric, enforced by both prompt design and test coverage (
backend/tests/test_tailoring.py).
- Security as a first-class concern, not an afterthought: HTTP Basic auth gates every API route, secrets are Fernet-encrypted at rest, Gmail access uses OAuth with a read-only scope, and CORS/env-var configuration was built specifically to support a genuine cross-domain production deployment rather than assuming same-origin dev conditions.
Screenshots
Add screenshots here after running the app locally/in a Codespace against the
sample data seeded by backend/db/seed_profile.sql, see docs/SETUP.md.
Design decisions worth knowing before you touch this
- No LinkedIn login automation, ever. LinkedIn coverage comes from (a) Gmail parsing of LinkedIn's own job-alert emails, and (b) JobSpy's unauthenticated public search (no account, so no ban risk, just IP-level rate limiting, capped conservatively).
- No autonomous submission in this pipeline. Every application is staged for manual review/approval in the dashboard. Full auto-apply across arbitrary company ATSs isn't reliably buildable and isn't the goal here: quality of application matters more than volume for these roles.
- CV tailoring never fabricates a skill.
services/tailoring.py can only reorder/select from skill_profile_items, see backend/tests/test_tailoring.py.
- No plaintext credentials, anywhere. Gmail access is OAuth-only (
gmail.readonly scope). Anything unavoidable is Fernet-encrypted with a key that lives outside the database.
- The API is never unauthenticated in a deployed environment. A single shared credential (
DASHBOARD_USERNAME/DASHBOARD_PASSWORD) gates every route except /health, checked with a timing-safe comparison, see backend/app/core/auth.py.
- GitHub Actions and Databricks are decoupled. Whether Databricks Free Edition supports external job-triggering is unconfirmed, so ingestion lands in Postgres on its own schedule and Databricks pulls from Postgres on its own native schedule, neither depends on calling the other.
Tech stack
| Layer | Choice |
|---|
| Backend | FastAPI, SQLAlchemy, Postgres (Neon, with pgvector) |
| Frontend | React, Vite, TypeScript |
| Pipeline | Python connectors + GitHub Actions cron |
| LLM | Groq (primary), Gemini (fallback), both free-tier |
| PDF generation | WeasyPrint |
| Deployment | Render (backend, Docker), Vercel (frontend) |
Repo layout
backend/: FastAPI + SQLAlchemy, serves the dashboard from Postgresfrontend/: React + Vite + TS dashboardpipeline/: ingestion connectors + GitHub Actions entrypoints + Databricks notebooksdocs/SETUP.md: one-time account/key setup (Neon, Google Cloud, Render, Vercel, etc.)
Running this yourself
See docs/SETUP.md for the full walkthrough. The short version: create a free Neon Postgres project, run backend/db/schema.sql then backend/db/seed_profile.sql (the fictional sample data above) against it, set the env vars in .env.example, and run the backend/frontend locally or in a Codespace.
About
Built by Megha Kanojia: LinkedIn · GitHub
