Data Systems / ML Systems · Fintech · production
Verifast
AI verification platform for income, identity, and fraud.
Production applicant verification system that orchestrates ID, bank-verified income, credit, criminal, and document checks into decision-ready workflows for multifamily leasing—with fraud detection and human-in-the-loop review.
Problem
Legacy screening rejects qualified renters on narrow income rules, while fraud and docu…
System
A parallel verification orchestration layer: bank + document income signals, custom OCR…
Role
AI Architecture & ML Systems
Status
production
Outcome
1.5M Renters verified / year
My role
AI Architecture & ML Systems
- AI Architecture
- Data Pipelines
- ML Systems
- Evaluation
- Productionization
Period / 2024–Present
Decision-ready applicant verification
Plays once · tap to replay
Problem
Legacy screening rejects qualified renters on narrow income rules, while fraud and document forgery force slow manual reviews and vacant units.
System
A parallel verification orchestration layer: bank + document income signals, custom OCR, fraud models, cross-checks to identity, and HITL experts—returning portfolio-configurable decisions into property management systems.
Architecture
Applicant data from PMS fans out into ID, bank, credit, criminal, and document/OCR paths; fraud models and a decision engine converge results, with optional HITL, before writing decision-ready outputs back to the PMS.
- PMS / Applicant → Verification Orchestrator
- Verification Orchestrator → ID / KYC
- Verification Orchestrator → Bank Income
- Verification Orchestrator → Doc OCR
- Verification Orchestrator → Credit / Criminal
- ID / KYC → Fraud Models
- Bank Income → Fraud Models
- Doc OCR → Fraud Models
Execution path plays once on view · hover a node or tap the canvas to replay
System anatomy
Architecture
Parallel verification fan-out with decision convergence.
- Orchestrator
- Partner checks
- Decision engine
- PMS write-back
Engineering decisions
Decision
Parallel verification fan-out
Constraint
Applicants need ID, income, credit, and criminal signals without serial bottlenecks.
Approach
Orchestrating checks in parallel and converging into one decision payload keeps time-to-lease short.
Result
Decision-ready outputs without Franken-processes across tools.
Decision
Bank truth + document OCR
Constraint
Document-only income is forgeable; bank-only misses some legitimate streams.
Approach
Combine instant bank connect with dual-model OCR and cross-checks to identity for full income reality.
Result
Higher approval of qualified renters without increasing fraud risk.
Decision
Custom OCR over off-the-shelf
Constraint
Generic OCR failed accuracy bars on messy applicant documents.
Approach
A dual-model extraction stack tuned for verification documents outperforms generic OCR in production testing.
Result
Reliable structured fields from diverse income paperwork.
Decision
HITL when models disagree
Constraint
Edge cases and fraud patterns still need expert judgment.
Approach
Same-day human-in-the-loop review (avg ~36 minutes) closes the loop without making every case manual.
Result
Automation when confident; humans when it matters.
Reliability
- SOC 2–aligned controlsACTIVE
- Fraud model monitoring / drift checksACTIVE
- HITL escalationACTIVE
- Data minimization and retention policyACTIVE
- PMS-native integrationsACTIVE
Evaluation
Outcome / Results
- Unified ID, bank income, credit, and criminal checks returned into PMS workflows.
- Support for W-2, gig, self-employed, benefits, and other nontraditional income.
- Fraud detection plus same-day human review when automation is uncertain.
Stack
Product visuals
