Resume Screening Pipeline

Gemini 2.5 Pro handles multi-column layouts and tables better than GPT-4o in 2026 benchmarks. Managed parsers (Affinda, Rchilli) are $0.01-$0.03 per resume and more accurate on straightforward layouts - use them for high volume and LLMs for edge cases.

Alternatives: Claude Sonnet 4 vision, GPT-4o vision, Affinda, Rchilli

Sonnet 4 follows structured extraction prompts reliably. Pinning extractions to a controlled taxonomy (ESCO for EU, O*NET for US) avoids the 'Python' vs 'python3' vs 'Py' explosion that breaks matching.

Alternatives: GPT-4o, Gemini 2.5 Pro, Custom NER with SkillNER

Presidio handles 30+ PII types out of the box and is extensible. Masking needs to go beyond standard PII: remove school names, graduation years, gendered pronouns, and photos. A stock PII service will miss several of these.

Alternatives: AWS Comprehend PII, spaCy + rules

Do NOT use an LLM as the final ranker. Regulators (NYC AEDT, EEOC) expect deterministic, auditable, replayable scoring logic. LLMs score resumes inconsistently on repeated runs of the same input, which is a compliance red flag.

Alternatives: Drools, Custom service

AIF360 has the widest coverage of fairness metrics (disparate impact, equal opportunity, demographic parity). You need quarterly bias audits documented per NYC AEDT and you need to demonstrate consistent scoring across protected classes.

Alternatives: Fairlearn, Holistic AI, Custom pipeline

Greenhouse Harvest is the best-documented ATS API and the most common in tech hiring. Inject scores as custom fields; do not replace the ATS's own pipeline stages. Recruiters keep control.

Alternatives: Lever API, Workday, SmartRecruiters

LLM ranker vs deterministic scorer

Using Claude or GPT-4o to directly rank candidates is tempting - they write great rationales. But LLMs are non-deterministic, score inconsistently on reruns, and cannot meet NYC AEDT's requirement of auditable, explainable, reproducible scoring. Use LLMs for extraction; use deterministic logic for scoring. This matters legally in the US and EU (AI Act classifies hiring AI as high-risk).

Managed parsers (Affinda, Rchilli) vs vision LLMs

Managed parsers are $0.01-$0.03 per resume and 95%+ accurate on standard layouts. Gemini 2.5 Pro vision is $0.15-$0.25 per resume but handles creative, multi-column, and image-heavy resumes. Route 80% of volume through a managed parser, 20% (designer/creative resumes) through the LLM.

Strict masking vs useful signal

If you mask too much (every school, every company), you lose signal a recruiter legitimately wants. If you mask too little, you leak bias. Compromise: mask names, photos, addresses, graduation years, and school names only for the first-pass score; reveal them at the recruiter-review stage after scoring is locked. Document and review this tradeoff with legal.

Resume parser hallucinates work experience or skills

Mitigation: Require every extracted fact to cite a span from the source text. Reject extractions without a citation span. Run a self-consistency check: re-parse 5-10% of resumes with a different model and flag disagreements above a threshold for human review.

Bias creeps in via proxy features (ZIP code, school, name-based gender/race signal)

Mitigation: Maintain a masked-resume baseline and run the scorer twice - once with full resume, once with masked. If scores diverge by more than 10% systemically for any protected class, freeze the model and investigate. AIF360 can detect disparate impact even when explicit protected attributes are absent.

Candidate appeals: 'why was I rejected?'

Mitigation: Log every per-requirement score with the source evidence text. Generate a human-readable rationale on request. Never auto-reject - require recruiter action to reject, and log the reason. NYC AEDT and GDPR Article 22 require a human-in-the-loop and an explanation.

Job description is itself biased (gendered language, unnecessary requirements)

Mitigation: Run the JD through a bias-check model (Textio, Ongig, or a custom classifier) before you use it as scoring criteria. Flag gendered terms, age proxies, and unnecessary years-of-experience requirements. Biased JDs produce biased matches regardless of how fair your screener is.

Model drifts after a promotion or firing decision affects training signal

Mitigation: Do NOT train ranking or scoring on 'which candidates got hired' - hiring decisions inherit past bias. Score against objective JD-match criteria only. Treat human override data as audit material, not training data. Re-audit quarterly.