Edge of Bliss Logo

Edge of Bliss AI Governance

Last Updated: September 23, 2025

Our promise: We use AI to open doors, not close them. We counter keyword-only filtering with adjacent skills and human judgment.

Contact: ai-governance@edgeofbliss.com | privacy@edgeofbliss.com | security@edgeofbliss.com 

Principles

  • Human-in-the-loop: AI suggests; people decide. No auto-reject/auto-advance.
  • Fairness by design: remove/clip proxy features; evaluate for disparate impact and performance parity.
  • Transparency: show why a match was suggested and how to improve it.
  • Privacy-first: minimal necessary inputs; purpose-bound use; opt-out from internal model training.

What Our AI Does (and Doesn’t)

  • Does: extract skills, map adjacent skills, summarize experience, and rank opportunities to widen the pool.
  • Does not: make legally significant decisions; use photos, names, or school pedigree in first-pass ranking.

How We Go Beyond Keyword Matching

  • Skill Graph: normalize titles, extract canonical skills, and expand to adjacent capabilities.
  • Context over counts: down-weight keyword stuffing; reward achievements, scale, outcomes.
  • Variance-tolerant parsing: evaluate across resume styles, languages, and formats.
  • Safety & robustness: red-team testing, drift monitoring, and clear incident playbooks.

Bias Mitigation Lifecycle

Design >Training > Evaluation > Deployment > Monitoring > Remediation

  • Design: exclude pedigree signals up front in firs-pass ranking; enable blind mode by default for screeners.
  • Data: no special-category demographics; internal model training only on de-identified/aggregated data; optional aggregate fairness checks if enabled by clients.
  • Training: balance datasets; adversarial tests (keyword stuffing, resume style variance).
  • Evaluation: relevance + fairness (impact ratios, recall purity, calibration; counterfactual and feature-sensitivity checks.
  • Deployment: guarded rollouts; human oversight; logging for audits.
  • Monitoring: drift alerts when monitored metrics shift>10% over 30 days; prompt blind mode and “expand my pool” where funnels narrow.
  • Remediation: re-weight features, remove proxies, raise human-review gates; document changes in the Model Card.

Explainability & User Controls

  • Why this match?: Show strongest factors (skills, experience bands, industry fit) and adjacent skills.
  • Blind mode: hide names, photos, and schools in first pass to reduce pedigree bias.
  • Expand my pool: One-click to include near-matches and adjacent-skills candidates.
  • Candidate clarity: “How to improve” suggestions (add missing skills, qquantify outcomes, link portfolio).

Governance & Monitoring

  • Decision role: Human-in-the-loop; no automatic accept/reject.
  • Owners: AI Governance Committee (Product + Trust & Safety + Legal + Eng).
  • Change management: Offline eval > bias checks > red-team > sign-off > limited rollout > live monitoring.
  • Incident classes: P1 for ranking visibility issues; PO if a proxy suppresses a protected cohort signal.
  • Vendors: third-party LLMs used for assistive features only; we do not train third-party foundation models on user content.
  • Privacy & security: PII-minimized prompts; encryption in transit/at rest; RBAC + MFA; audit logs.

Regulatory Alignment (Snapshot)

Framework What we commit to Status / Key date
NYC Local Law 144 (AEDT) Independent bias audit by employers who use tools as AEDT; publish summary; candidate notices. Edge of Bliss provides audit hooks and reports. Required before use in NYC.
California FEHA (ADS regulations) Support recordkeeping (≥ 4 years), reasonable-accommodation workflows, and configuration guidance to avoid discriminatory use. Effective Oct 1, 2025.
Colorado AI Act Map to “reasonable care” duties; maintain a risk management program; provide deployer notices; support post-market monitoring. Delayed to Jun 30, 2026.
EU AI Act (employment = high-risk) Risk management, data governance, logging, human oversight, and post-market monitoring; publish Model Cards and maintain conformity evidence. High-risk obligations phase in on EU timeline.
Illinois AI Video Interview Act Provide notices/consents, limit sharing, and honor deletion requests where employers use video analytics; Edge of Bliss provides hooks and guidance. In force.
EEOC Title VII guidance Assess adverse impact; apply UGESP concepts to AI-assisted selection; maintain testing documentation. Ongoing compliance.

We review this table regularly and update dates/requirements as regulations evolve.

Model Cards & Auditability

For each major AI capability we maintain a Model Card covering: purpose & limits, inputs used/excluded, evaluation results (including fairness metrics where applicable), known risks, monitoring plan, and change history.

Contact the AI Governance Team