dsh-resume-screening

Receiving hundreds of resumes daily yet having no time to review each one is the reality for many HR professionals. Throwing resumes straight at an LLM for screening quickly exhausts the context window, costs scale linearly with volume, and rigid criteria are unstable when left to a model to interpret. dsh-resume-screening takes a straightforward approach: ingest resumes in bulk, build a structured archive, and run screening inside the database. Only the short-listed candidates pass through the LLM for fine-grading. A single repository handles up to 100,000 records across docx, pdf, xlsx, md, and txt formats, with content-hash deduplication so repeated submissions never create duplicates.

The workflow has three stages. Ingestion converts every resume into a unified Markdown archive; a rule engine extracts hard fields such as education, tenure, and school first, then an LLM pass fills in fine-grained skills and semantic categories, all tagged with confidence scores and provenance so the profile is built once and reused forever. Screening takes natural-language job criteria (mandatory hard filters plus optional weighted bonuses), narrows the pool at millisecond speed in the database, then reads each finalist in full and returns a 0-to-100 fit score with a recommendation tier. Saved job templates keep the standard identical across every new batch.

Ideally suited for small-to-mid HR teams, independent hiring managers, and any workflow that repeatedly screens against the same criteria. The engine uses SQLite as a queryable index and Markdown as the single source of truth, with a full-rebuild path available at any time. LLM calls reuse the host model with the reasoning chain explicitly disabled for structured requests, keeping both latency and cost in check.