IntelliTalent - Resume to Job Matching
End-to-end resume-to-job matching platform using PyMuPDF parsing, GPT-4o-mini skill extraction over a FAISS-backed RAG pipeline, and hybrid 60% cosine / 40% skill-overlap scoring on live listings.
Overview
An NLP-driven platform that matches resumes to live job listings. Resumes are parsed with PyMuPDF, skills extracted using GPT-4o-mini via a FAISS-backed RAG pipeline, and jobs ranked with a hybrid score of 60% cosine similarity and 40% skill overlap over listings scraped from Apify. Latency is optimized with regex taxonomy matching (50ms vs. 5+ seconds per-job LLM calls) and batched OpenAI embeddings with vectorized NumPy cosine similarity. Deployed across AWS EC2 and GCP Cloud Run with O*NET role clustering and MongoDB text search.
Problem Statement & Approach
Job seekers waste time manually scanning listings that turn out to be irrelevant. Keyword-based search produces both false negatives and noise, and gives no clarity on how well a resume actually qualifies for a role.
Approach: Parse the PDF resume with PyMuPDF, extract structured skills with GPT-4o-mini over a FAISS-backed RAG index, scrape live listings from LinkedIn and Indeed via Apify, embed everything with OpenAI's text-embedding-3-small, and rank with a hybrid score of 60% semantic similarity plus 40% skill overlap — returning ranked matches with skill-gap analysis.
System Architecture
A multi-stage pipeline moves a resume from upload to ranked matches: PDF upload → PyMuPDF extraction → GPT-4o-mini skill extraction via FAISS RAG → Apify job scraping → batch embeddings → vectorized cosine similarity → hybrid scoring → MongoDB storage → a web dashboard of ranked matches.
Key Features
- FAISS-backed RAG for GPT-4o-mini skill extraction
- Hybrid job ranking: 60% semantic similarity + 40% skill overlap
- Regex taxonomy matching replaces per-job LLM calls (~50ms vs. 5+ sec per job)
- Batch OpenAI embeddings with vectorized cosine similarity
- O*NET role clustering with GPT-generated role suggestions
- Bcrypt auth with OTP verification and brute-force protection
- MongoDB TTL indexes auto-expire scraped listings after 7 days
Technical Stack
LLM & Embeddings
Core
Data
Deployment
Security
Deployment
Containerized with Docker and deployed through AWS (deploy_aws.sh) and GCP Cloud Run (deploy_gcp.sh) scripts, with secrets managed via AWS SSM Parameter Store and GCP Secret Manager.
Challenges & Solutions
Challenge: Per-job LLM calls made ranking too slow — 5+ seconds per listing.
Solution: Regex taxonomy matching replaced per-job LLM calls, cutting skill matching to ~50ms per job.
Challenge: Feeding the full skill taxonomy to the model added heavy token overhead.
Solution: FAISS-injected RAG context supplies only the relevant taxonomy slice per resume.
Challenge: Scraped job listings went stale and bloated storage over time.
Solution: MongoDB TTL indexes automatically expire listings after 7 days.
Improvements
- Application tracker to manage jobs a user has applied to
- Resume quality scoring with actionable feedback
- Cross-user alerts when strong new matches appear
- A personal analytics dashboard for search activity
