AI-Powered Resume Screener
Automating recruitment workflows for an HR tech startup using advanced NLP and LLM embeddings.
The Problem
An HR tech startup serving enterprise clients needed to screen 10,000+ resumes daily across multiple job roles. Their manual process consumed 40+ hours per batch, and inconsistent evaluations between reviewers led to high false-negative rates — qualified candidates slipped through while underqualified ones advanced.
The Architecture
We designed a two-stage AI pipeline using Python, FastAPI, and OpenAI embeddings. The AI integration approach prioritised accuracy over speed, with each stage serving a distinct purpose.
Stage 1: Semantic Retrieval
Resumes were parsed and chunked into 512-token segments, then embedded using OpenAI text-embedding-3-large (3072 dimensions). Pinecone vector database performed similarity search against structured job descriptions, returning the top 20 candidates from a pool of thousands — in under 200ms.
Stage 2: LLM Evaluation
The top 20 candidates from Stage 1 were sent to GPT-4 for structured deep evaluation. The model output a scored JSON object with strengths, weaknesses, and a confidence score per job requirement. Candidates below a configurable threshold were automatically flagged for human review, ensuring no critical hire was made on AI judgment alone.
Measured Results
- 90% reduction in screening time: from 40 hours to 4 hours per 10k resumes
- 95% accuracy against human-reviewed benchmarks across 3 months of parallel runs
- ₹12L/month saved in recruiter operational costs
- 3x faster time-to-hire across all client positions
The system processes resumes in English, Hindi, and Marathi with equal accuracy, and has been extended to screen for behavioural fit using the same embedding pipeline.
Explore related: AI integration services, Backend architecture projects.