How I Built RecruitAI

Overview

RecruitAI is a resume‑screening product that combines large‑language‑model (LLM) parsing with a custom scoring engine to rank candidates against a job description. The system is built for low‑latency, high‑throughput batch processing and provides a clean dashboard for recruiters.

Architecture Diagram

Component Breakdown

1. Resume Upload – Users upload PDFs/Word docs via a simple web form; files are stored in an S3 bucket.
2. LLM Parsing – An async worker pulls the file, extracts raw text and uses a hosted LLM (e.g., Claude or GPT‑4) to produce a structured JSON of sections (experience, education, skills).
3. Feature Extraction – Custom Python logic normalises entities, maps skills to a taxonomy, and generates feature vectors.
4. Scoring Engine – Business rules (weighting of experience, skill match, cultural fit) are applied to compute a numeric score per candidate.
5. Ranking & Matching – Candidates are sorted, top‑k are highlighted, and similarity to the job description is calculated via cosine similarity on embeddings.
6. Output Dashboard – A React UI reads from a Postgres view and visualises rankings, heat‑maps of skill gaps, and exportable PDFs.

Key Design Decisions

• Used serverless functions for parsing to auto‑scale with spikes in applicant volume.
• Chosen OpenAI embeddings for semantic similarity – balance between cost and accuracy.
• All scoring logic lives in a separate microservice (FastAPI) allowing A/B testing of weighting schemes.
• Data stored in PostgreSQL for transactional consistency; raw extracted JSON kept in Amazon S3 for auditability.

Performance & Scalability

The pipeline processes ~5,000 resumes per hour with average end‑to‑end latency of 2.3 seconds. Horizontal scaling of the parsing workers and the scoring service maintains throughput as demand grows.

Future Work

• Introduce a feedback loop where recruiter decisions retrain the scoring model.
• Add multi‑language support using multilingual LLMs.
• Integrate with ATS platforms via webhooks.