Signal Review
When an application queue grows into the thousands, speed matters. The first pass also needs criteria a reviewer can inspect, source evidence, and an explicit place for uncertainty.
More applications than a recruiter can read in one sitting.
A generic keyword filter misses wording variants; an opaque score gives little help when someone asks why a résumé was surfaced. An AI model can be persuasive even when its explanation is detached from the source. A useful first pass needs to retrieve job-relevant evidence quickly while keeping a human responsible for review.
Rank by stated criteria. Show the sentence.
The React workbench generates 3,518 fictional candidate profiles or loads a local CSV, parses a weighted rubric, and sends the batch to a Web Worker so the page stays responsive. Each found criterion carries the exact résumé sentence; missing required evidence marks the profile for review rather than rejecting it. The local Python pipeline ingests text, PDF, and DOCX files into SQLite and records a run with source digests, rubric, scores, and evidence.
A local Node benchmark evaluated 3,518 synthetic profiles against eight criteria in 421 ms on one machine. This measures deterministic evaluation only, not file parsing, model latency, network, UI rendering, or human work. The live page reports its own measured worker time on your device.
- WORKBENCH
- React · TypeScript · Web Worker · CSS
- LOCAL PIPELINE
- Python · SQLite · PDF/DOCX extractors
- VERIFICATION
- Source quotes · unit tests · reproducible benchmark
Model output needs a source check.
The hosted demo uses a deterministic, browser-local evidence engine and does not call a model. The optional local model path sends only a bounded top subset of résumé texts to a structured extraction API when an operator enables it. Each returned quote is checked as an exact substring of the source; an invalid quote becomes unknown, and the score is recalculated from validated evidence. The model cannot issue a rejection.
The lexical first pass can miss strong candidates with different wording. Before production use, this needs recall evaluation, bias analysis, a broader retrieval layer, access controls, retention rules, reviewer override and appeal workflows, and a real integration test against an applicant system.
Run the queue. Open a candidate. Inspect the proof.
Change the job criteria, run the synthetic batch, search the review queue, or import a local CSV. Imported data stays in the browser tab. The generated applications are for exercising the interface and throughput; they are not real applicants or a benchmark for hiring quality. This project is a technical reference, not a deployed ATS or a hiring decision service.
REFERENCE SYSTEM · SYNTHETIC PROFILES · HUMAN REVIEW REQUIRED