Built a lightweight but production-grade AI resume screening engine — the core intelligence layer extracted from OmniATS, open-sourced without any proprietary multi-tenancy or billing code.

The problem it solves:

Manual resume screening is the biggest bottleneck in any hiring pipeline. A recruiter spending 6 minutes per resume across 200 applicants is 20 hours of cognitive work before a single interview is scheduled. This engine does it in seconds with structured, explainable output.

How it works:

→ Takes two inputs: a job description (text) and a candidate resume (PDF) → Extracts structured data from the resume using Claude API — skills, experience, education, and key achievements as typed fields → Scores the candidate against the job description on a 0–100 match scale with field-level reasoning → Returns a Pydantic-validated output object — candidate name, match score, extracted skills, experience summary, and recommended action → Auto-correcting feedback loop: if Claude returns malformed JSON, the validation error is injected back automatically for self-correction → Zero frontend — pure Python CLI and API, designed to be embedded into any existing hiring workflow or ATS backend

What makes this production-safe:

Every output is schema-enforced via Pydantic v2. The LLM cannot return an unstructured response that breaks downstream logic. The scoring is deterministic in structure even when the reasoning varies — you always get a float, a list, and a string. Never a surprise.

Stack: Python · Anthropic Claude API · Pydantic v2 · PyMuPDF · Flask

Open source → https://github.com/gharisj3/llm-resume-screener