AI Transparency
We believe you deserve to know how decisions about your career are made. Here is a complete breakdown of our AI systems.
We match candidates to jobs using a weighted multi-factor system. Every match score is explainable.
How closely your skills and certifications overlap with what each role requires.
How relevant your work history, roles, and years of experience are to the position.
Whether the job's location, remote options, and salary range fit your stated preferences.
These factors combine under a fixed, published two-sided formula: a candidate-fit score (skills 40%, experience 20%, verified evidence 25%, recent activity 15%) and a job-appeal score (salary fit 30%, location 25%, employer quality 25%, salary transparency 10%, freshness 10%), joined by harmonic mean so one-sided matches rank low, optionally blended with semantic text similarity. The score is always shown alongside the full job listing so you can judge for yourself.
Looking for the full AI transparency policy — the data our AI does and does not use, your right to an explanation, opt-out rights, and how to contact our Data Protection Officer?
Read the AI transparency policyWe actively monitor and prevent discriminatory outcomes in our AI systems.
Selection rates for any protected group should reach at least 80% of the rate for the highest group — the fairness target we design our matching toward.
We are building a recurring internal review process to measure disparate impact across gender, age, ethnicity, and disability status, and aim to publish results as the program matures.
We use general-purpose language models rather than training our own, and design our prompts and ranking signals to reduce cultural bias across regions and industries.
When AI confidence is low or outcomes appear anomalous, human reviewers intervene before any decision is finalized.
All AI algorithms used by ZhiYin are registered and documented, compliant with China CAC and EU AI Act requirements.
7 algorithms documented in our transparency registry, designed to align with CAC and EU AI Act requirements
| ID | Algorithm |
|---|---|
| ALG-001 | Job-Candidate Matching (two-sided reciprocal) Rank candidate–job pairs with a deterministic, fixed-weight TWO-SIDED formula: a candidateFit score (skills 40% via bilingual EN-ZH lexicon, experience 20%, verified evidence 25% — mock-interview/proctored-assessment/course/resume signals weighted by published predictive validity — and activity recency 15%) and a jobAppeal score (salary fit 30%, currency-gated; location 25%; employer quality 25% from reply rate, offer-acceptance rate and candidate satisfaction; salary transparency 10%; job freshness 10%), combined by harmonic mean so one-sided matches rank low, optionally blended 60/40 with multilingual text-embedding similarity. Feed ordering additionally applies aggregate-level congestion balancing (over-applied jobs demoted up to 12%, under-exposed jobs boosted up to 6%). Every score ships with a per-component breakdown and matched/missing-skill evidence; no LLM participates in scoring; protected characteristics and hukou are never inputs. |
| ALG-002 | Resume Analysis Extract and score skills, experience, and qualifications from resumes |
| ALG-003 | Cover Letter Generation Generate personalized cover letters based on candidate profile and job description |
| ALG-004 | Interview Question Generation Generate relevant interview questions based on job role and career track |
| ALG-005 | Salary Insights Provide salary range estimates based on role, location, and experience level |
| ALG-006 | AI Team Composition (team-compose-v1) Compose 3 ranked contractor teams from an employer brief by scoring candidates on skill overlap, rate fit, verification, availability, and past rating, then generating per-member rationale via LLM. |
| ALG-007 | Job Search Ranking Order public job-search results with a deterministic blend when a text query is present: text relevance (Postgres full-text rank or weighted keyword hits with bilingual alias and Chinese word-segmentation expansion) 55%, freshness (30-day exponential decay) 25%, and aggregate engagement (recency-weighted views and application conversion, computed daily, no per-user profiling) 20%. Employer responsiveness tier remains the primary ordering; chronically unresponsive employers rank below all others. No LLM participates; no individual user behavior is used. |
Our AI systems are designed to align with employment-discrimination laws and AI-governance regulations in the markets we serve.
We design our automated decision tools to support bias review in line with emerging US automated-employment-decision-tool rules.
Recruitment AI is treated as high-risk under the EU AI Act, and we are building the Article 9 bias-monitoring and human-oversight practices it calls for.
Our recommendation and matching algorithms are designed to support the algorithm-registration requirements of the Cyberspace Administration of China.
Join a platform where AI works for you, not against you.
Get Started Free