Six agents. Four minutes. Zero PII in the cloud.
Process mortgage applications in minutes, not days. Shift-left detection, semantic anonymization, and six specialized AI agents — all on your premises.
THE PIPELINE
Six Agents, One Application
Each agent is specialized. Documents flow through intake, extraction, validation, analysis, compliance, and delivery — with anonymization boundaries protecting PII at every step.
Iris
Intake
Classifies uploaded documents — W-2, 1099, bank statements, tax returns, appraisals, IDs. Routes each to the right extraction path.
Cost: $0Rex
OCR Extraction
Multi-engine cascade: GLM-OCR → MiniCPM → Qwen2-VL → Tesseract. Each model tries in order until confidence threshold is met.
Cost: $0Val
Validation
Shift-left deterministic checks — SSN conflicts, name variant mismatches, table math errors, missing documents. Zero AI cost.
Cost: $0Ana
Analysis
Local LLM (Qwen 35B) detects anomalies — phonetic mismatches, employment gaps, income inconsistencies. Runs on your GPU with full PII access.
Cost: LowClaire
Compliance
Claude checks 15+ Fannie Mae/FHA/VA rules on anonymized data only. DTI/LTV ratios, documentation requirements, regulatory compliance.
Cost: ~$0.04Max
Delivery
Deanonymizes the compliance report, pushes to your LOS (Encompass API), triggers voice callback, purges temporary mappings.
Cost: $0FEATURES
Everything You Need for Autonomous Underwriting
Built for mortgage professionals who want AI speed without sacrificing compliance, privacy, or control.
Shift-Left Detection
Val catches SSN conflicts, name mismatches, and math errors deterministically before any LLM runs. Zero cost, instant results.
Semantic Anonymization
PII is stripped before cloud processing. Claude never sees real names, SSNs, or addresses. Mappings are ephemeral and purged after delivery.
Multi-Engine OCR
Four-model cascade (GLM-OCR, MiniCPM, Qwen2-VL, Tesseract) extracts data from any document quality. Confidence scoring picks the best result.
Tiered AI Architecture
Deterministic checks are free. Local LLM handles sensitive data. Cloud AI only sees anonymized content. Each tier optimized for cost and privacy.
15+ Compliance Rules
DTI/LTV ratios, employment verification, income documentation, appraisal guidelines — Fannie Mae, FHA, and VA rules built in.
Pixel Office Dashboard
HTMX-powered dashboard with animated agent characters, CRT terminal with live logs, and real-time pipeline state streaming.
OCR Lab
Inspect extracted data field-by-field. Compare OCR engine outputs. Debug extraction accuracy with confidence scores and bounding boxes.
Nancy Voice Assistant
Natural language commands for the underwriting pipeline. Ask Nancy to process an application, check status, or explain an anomaly.
Anomaly Intelligence
Three detection tiers: deterministic (free), local LLM (private), cloud LLM (anonymized). Phonetic analysis, gap detection, cross-document validation.
Adaptive Model Selection
Simple cases use Haiku. Complex cases escalate to Sonnet or Opus. Privacy-sensitive operations stay on local Qwen 35B. Up to 60% cost savings.
Fault-Tolerant Pipeline
Exponential backoff with jitter. Automatic retries on transient failures. Fallback agents for critical stages. 99.5% completion rate target.
LOS Integration Ready
Encompass API integration for production deployment. Deanonymized reports push directly to your loan origination system.
PRIVACY ARCHITECTURE
PII Never Leaves Your Premises
The anonymization boundary between Ana and Claire ensures that real personally identifiable information is never sent to cloud AI. Only anonymized tokens cross the wire.
123-45-6789 → SSN_A
Mapping purged
Ephemeral Mappings
Anonymization mappings exist only in memory during processing. They are never written to disk and are purged immediately after delivery.
Local-First Processing
Iris, Rex, Val, and Ana run entirely on your hardware with full PII access. Only Claire (compliance) uses cloud AI — on anonymized data.
Audit Trail
Every step is logged with timestamps. The audit records that anonymization was applied, without storing the actual mapping.
HOW IT WORKS
See the Pipeline in Action
FAQ
Frequently Asked Questions
How does the anonymization work?
Before any data reaches cloud AI (Claire), the Ana agent creates an ephemeral mapping that replaces all PII with tokens (e.g., "John Smith" becomes "APPLICANT_A"). Claire makes compliance decisions on anonymized data. After processing, the mapping is reversed for the final report, then permanently destroyed. The mapping never touches disk.
What documents can it process?
W-2s, 1099s (all types), bank statements, tax returns (1040 with schedules), pay stubs, appraisal reports, employment verification letters, gift letters, purchase agreements, title reports, and government IDs. The multi-engine OCR cascade handles varying document quality.
What compliance rules are built in?
DTI front-end and back-end ratio checks, LTV validation by loan type, credit score minimums, employment history continuity, income and asset documentation completeness, gift letter requirements, appraisal validation, comparable sales analysis, flood zone determination, and occupancy type verification — covering Fannie Mae, FHA, and VA guidelines.
Can it integrate with our LOS (Encompass, etc.)?
The Max delivery agent is designed to push deanonymized compliance reports to Encompass and other loan origination systems via their APIs. The integration module supports configurable endpoints and authentication.
What happens when an anomaly is detected?
Anomalies are tiered: Tier 1 (deterministic) catches SSN conflicts and math errors instantly at zero cost. Tier 2 (local LLM) finds phonetic mismatches and employment gaps using Qwen 35B on your hardware. Tier 3 (cloud) handles complex regulatory analysis on anonymized data. Each anomaly gets a severity rating and detailed explanation.
How much does it cost per application?
The demo shows an average of $0.04 per application in AI costs. Four of the six agents run at zero cost (deterministic or local GPU). Only Claire uses cloud AI, and only on anonymized data. The adaptive model selector can further reduce costs by routing simple cases to cheaper models.
Can we run it entirely on-premises?
Yes. The pipeline is designed for on-prem deployment. Iris, Rex, Val, Ana, and Max run entirely on your hardware. Claire can be configured to use a local LLM (Qwen 35B) instead of cloud Claude, making the entire pipeline air-gapped if required.
What about the shift-left approach?
Val (the validator agent) catches critical issues — SSN conflicts, name mismatches, missing documents, math errors — using pure deterministic logic before any LLM is involved. This means obvious problems are caught instantly at zero cost, and the more expensive AI agents only process applications that pass basic validation.
GET STARTED
See Mortgage Lite in Action
Book a demo to see the full pipeline process real mortgage documents in under 5 minutes.
Ajay Tyagi, VP Sales — ajay@dkube.io
US Dkube Contact: +(1) 408 430 2503