LOAN FILE
BORROWER(S)--
PROPERTY VALUE--
LOAN AMOUNT / LTV--
STATED FICO--
CHANNEL--
DECISION STATUS--
Underwriter actions
Total Micro-Tasks
0
Executed on current file
Discrepancies Flagged
0
Requiring human underwriter action
Passed / Resolved
0
Cleared conditions
Audit Trail Records
0
Immutable forensic events
Outstanding Documents
0
Partner instruments still needed
Outstanding Documents Ask-for inventory — not audit failures
Loading document inventory…
Actionable Exceptions Flagged discrepancies only — open Check Audit for full analysis & evidence
Filter by Tag
Loading task queue from DuckDB...
Select a check

Choose a micro-task from the list to review analysis, audit steps, and screenshots.

Architecture

How Fusion Micro-Tasking Works

The core evaluation engine and audit checks are 100% deterministic code. AI is used strictly at the perimeter for natural-language interaction and semantic classification — never to decide APR tolerances, escrow math, or stare-and-compare matches.

Deterministic code AI / LLM assist Underwriter review
Document processing data flow Where code runs vs where AI assists

From package drop to cockpit exceptions: each stage shows whether the work is code, AI, or human.

1
CodeAI chat

Package intake

PDF lands in inbox/ (chat attach or drop). Orchestrator inbox_process_package promotes it to loans/<id>/package/.

2
Code

Page index & OCR

Python indexes pages (package_page_index), extracts text, and maps Closing Disclosure / Note / flood pages. OCR runs only when vector text is missing.

3
Code firstAI fallback

Document classification

Filename + text heuristics pick Credit / Appraisal / Assets / Closing. Ambiguous headers can call a lightweight classifier (~100 ms) — not the audit decision.

4
Code

Visual evidence render

Sandbox rasterizes PDF pages, draws catalog-anchored highlight boxes, writes PNG + *_thumb.jpg via /tmp/output/, then promotes to loans/<id>/evidence/.

5
Code

Highlight QA gate

validate_highlight_box.py rejects floating boxes, column bleed, or neighbor-label intrusion (exit code 2). Failed proofs are not attached.

6
Code

Deterministic micro-tasks

Suite wrenches run CD / CR / FL / AP / AS checks: TRID math, RESPA escrow cushions, FICO mid-score, stare-and-compare strings & amounts — all JavaScript/TypeScript + SQL.

7
Code

Immutable audit trail

Results land in DuckDB (loan_task_instances, task_audit_log) with severity rollup. Evidence paths point only at durable loan storage.

8
HumanAI explain

Cockpit & exception review

This dashboard shows flagged discrepancies + proofs. The assistant can explain citations and draft condition language — it does not re-score the math.

100% deterministic code Zero AI hallucination on decisions
  • 46+ audit checks CD-001–036, FL-001–004, CR-001–006 — TRID tolerances, APR variance, 3-business-day timing, RESPA §10 escrow cushion, stare-and-compare LOS vs document fields, tri-merge mid-score & liability matching.
  • Visual proof pipeline PDF rasterization, page discovery, coordinate extraction, tight field highlights (draw_field_highlight.py), automated QA gate with programmatic exit codes.
  • Persistence & rollup Storage promotion, dedupe, immutable task_audit_log events, and loan decision severity via DuckDB views.
Where AI is still used Perimeter only
  • Semantic classification Lightweight classifiers for ambiguous document headers or credit inquiry names (e.g. auto vs mortgage inquiry) when rules are inconclusive.
  • Conversational cockpit Underwriter Q&A, exception narration, regulatory citation explanation, and draft condition letters — grounded in stored check results.
  • OCR text fallback Degraded scans without native PDF text vectors; extracted strings still feed the same deterministic checks.
Why it was built this way

In mortgage underwriting and compliance audits, rules engines cannot rely on probabilistic LLM reasoning to decide whether an APR is out of tolerance or a loan amount matches. Fusion uses deterministic code for every math check, tolerance test, and visual receipt — and reserves AI for document comprehension, categorization, and conversational assistance.

Code
Pass / fail / severity
Code
Evidence geometry & QA
AI
Classify & converse
Human
Exception disposition
Lender Micro-Task Catalog Modular underwriting rule suite

Choose a channel, then enable the micro-tasks that should run for that channel. Batch select to enable, disable, or tag rules without leaving the catalog.

Editing channel Enablement below applies only to Retail
Suite
Status
Tags