Explainable recommendations. Human-in-the-loop by design.
Every HSI recommendation ships with supporting evidence, risk flags, reason codes, model card, and a fairness audit — so landlords can defend the decision under FCRA scrutiny and internal governance review.
- 12 mo · rent on-time
- RTI 27% (band OK)
- 3 income sources verified
- Thin credit file
- Documented 2023 hardship
- Protected-class inputsNone
- Approval-rate parity checkWithin band
- Housing-type penaltyDisabled
- Human review requiredYes
The explainability layer (audit trail, reason codes, decision rationale panel, and model versioning) is live in production today as part of the HSI landlord workspace. The broader AI Decision Support product — with fairness monitoring dashboards, cohort audits, and bias-detection pipelines — is in development.
Capabilities.
Recommendation + confidence
Every applicant surfaces a plain-English recommendation (Approve / Conditional / Cosigner / Decline) plus a model confidence value.
Supporting evidence + risk flags
Two-column panel: positive evidence supporting approval, plus explicit risk flags reviewers must consider before decisioning.
Reason codes
Machine-readable reason codes on every decision — the exact tokens that appear on adverse-action reports.
Fairness audit
Every decision includes a fairness posture check: protected-class inputs status, approval-rate parity, housing-type penalty status, and human-review requirement.
Model card + versioning
Model version, release date, weight configuration, and input hash stamped on every decision — trivially reproducible for audit.
Bias monitoring dashboard
Portfolio-level view of approval, decline, and conditional rates across proxy dimensions — with alerting when parity bands drift.
Built for enterprise operators.
Not a black box
Every input, every weight, every model version, and every decision reason is surfaced — no hidden features, no opaque outputs.
Human retains final authority
AI produces recommendations. Landlords produce decisions. No autonomous adverse action — ever.
Reason codes match adverse-action doc
The tokens shown in the AI panel are the same ones that appear on the FCRA-style report — no drift between UI and legal document.
Fair-housing aware from day one
No protected-class variables. Housing type is never a scoring input. Recovery patterns are rewarded, not penalized.
Where this is going.
Dates are directional. Enterprise pilot partners shape the sequencing.
- Explainability panel live
- Reason codes on adverse-action
- Model card v1.1.0
- Structured audit logging
- Portfolio-level fairness dashboard
- Approval-rate parity alerting
- Model card publication for v1.2
- External bias-audit review
- Cohort audit exports
- Public model card page
- Fairness posture API
- Third-party model attestation
Run this against your live portfolio.
Enterprise pilots include a backtest on your historical applicants, a signed DPA, and a dedicated integration engineer.