jev-resume-screening
Resume-to-JD screening study (Chinese) with 5 Noul gates, 4 Score dimensions and a routing Choice per request, hardened over three versions.
Overview
This project, documented in Chinese, packages a resume-to-job-description screening design with Jev for a user-operations role. A structured state holds the decomposed job description and the PII-stripped resume, and one request asks ten parallel questions: five Noul evidence gates (education threshold, written English evidence, heavy AI product use, efficiency tools at work, exceptional pursuits), four 0-3 Score dimensions (community work depth, data and metrics ability, cross-team delivery, signal-to-insight), and one Choice routing the candidate's dominant background to the right recruiter. There is deliberately no overall match question; code computes totals.
The author documents the process: reading the docs, a failed coin-flip test that led to rules (have material, ask one judgment, define criteria clearly), Chinese and English versions of a test resume, revisions, and negative-control resumes. Findings include healthy calibration (factual questions near 0.99, judgment questions lower), and that confidence drifted across languages (0.67 Chinese versus 0.75 English), which would change routing at a 0.7 threshold. In v3 a glossy-trap resume's AI-heavy-user answer fell from 0.95 to 0.49. Low-confidence answers go to human review.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the resume show evidence of a bachelor's degree or above? | yes · no |
| Score | How deep is the candidate's user-community work (0-3)? | 0 · 1 · 2 · 3 |
| Choice | dominant_background: which background bucket fits the candidate? | — |
How it works
- Decompose the job description into structured state
- Strip PII from the resume
- Ask 10 questions in one parallel request
- Compute totals in code
- Escalate low-confidence answers to human review
Strengths
- Detailed, honest experiment log
- Negative-control testing of criteria
- Shows cross-language threshold drift
Caveats
- Single role and synthetic resumes
- Mostly design documents and results
- Chinese-language documentation
At a glance
- Type
- Project
- Status
- research
- Runtime
- Criteria templates and test requests
- Has code
- yes
Use cases and integrations
How curated lists describe it
Recruiting: screens one resume against a JD in a single request of five Noul evidence gates, four Score dimensions, and one background-routing Choice, with criteria hardened v1→v3 against negative-control resumes (a glossy-trap CV's self-described "AI heavy user" fell 0.95→0.49) and any low-confidence answer escalated to human review.
yibie/awesome-jevThis project integrates Jev to provide structured decisions for its workflow. See the repository for implementation details.
logicrw/awesome-jev-projects
Frequently asked questions
Why no overall score question?
Totals are computed in code from atomic answers.
What risk did testing reveal?
Confidence differed between Chinese and English versions of the same resume, which could flip routing at a fixed threshold.
How were criteria improved?
Through v1-v3 revisions tested against negative-control resumes.
Related projects
CLI that has Jev judge a video's subtitles segment by segment, then cuts a highlight reel and a filler-free full version.
ProjectbetaPythonPython library that replaces hard-coded RAG thresholds with calibrated decisions for stopping, chunking, selection, abstention and caching.
ServicebetaPythonResume review app and MCP server where a Cloudflare Worker asks Jev typed questions and composes a JevScore, generally or per job.
AppbetaTypeScriptDaily job-search pipeline that triages and scores public postings against a resume with Jev, then applies via a Grok Bot webhook.
ServiceprototypeTypeScriptCV review app where Jev makes traceable document-level judgments on clarity, contribution and results, optionally against a job posting.
ProjectprototypePythonLocal CV screening workbench that asks Jev fixed typed questions per CV and scores them with editable, free-to-rerun policy arithmetic.
Projectprototype