Case studyOrigin
Every loan document,
understood.
Origin is an AI loan-origination platform. Loan officers create and process entire loans while the AI ingests any document — however messy — classifies it, extracts the data, and moves the file forward. With a social layer where finance pros post, comment, and share.
Pipeline active
W-2 · 2025→ extracted
Bank statement→ extracted
Pay stub→ extracted
Scan · page 7review
1040 tax return→ extracted
Driver's license→ extracted
processed today 4,820accuracy 99.2%
# EngagementLoan-origination platform
# RoleDesign → deploy
# StackReact · Python · Cloud Run
# ModelsGemini Flash · Qwen
# ToolingOpenClaw · vision OCR
# SurfaceMulti-tenant + social
# 01 /The problem
Loan files drown
in documents.
Mortgage processing is paperwork at scale — every borrower's packet is different, every scan a little messy. Sorting, classifying, and keying it by hand is slow, expensive, and error-prone. The loan stalls while the documents pile up.
Origin turns the document pile into a processed loan file.
# 02 /What we built
Upload to originated.
One pipeline.
Any document a borrower sends becomes structured, validated loan data — classified, extracted, and routed into the right file without manual sorting.
01UploadAny doc, any format
02ClassifyDoc type detected
03ExtractFields + tables
04ValidateRules + review
05Loan fileStructured record
06OriginateLO workflow
# 03 /How it works
Any document in. Structured data out.
A multi-model pipeline reads the document the way a processor would — but in seconds. It names the type, pulls every field, and scores its own confidence so reviewers only touch what matters.
- Classifies and splits combined packets automatically.
- Extracts fields, tables, and totals with confidence scores.
- Routes low-confidence fields to human review, not the whole doc.
# 04 /Architecture
Four layers.
Multi-model, multi-tenant.
A React workspace on top, a Python processing API on Cloud Run, a multi-model AI core, and a multi-tenant platform with roles and the social graph.
L1WorkspaceReact · Next.jsLO console, borrower document workspace, social feed, tenant admin.
L2ProcessingPython · Google Cloud RunIngestion, classification, extraction, validation — autoscaled per tenant.
L3AI modelsGemini Flash · Qwen · OpenClawFast classification, structured field extraction, and vision OCR for complex scans.
L4PlatformMulti-tenant · roles · socialLO, processor, and tenant-admin permissions; audit trail; the community graph.
# 05 /Results
The numbers
behind the file.
What AI document processing looks like once it's running inside a real loan-origination workflow.
0+
document types classified automatically — from W-2s to the messiest bank statements.
0.0%
field-extraction accuracy across validated production documents.
0×
faster than manual processing — a full loan package handled in minutes, not days.
0
manual sorting — every page routes itself into the right loan file.
↳ Illustrative figures — replace with audited numbers before publishing.
# 06 /Capabilities
What Origin ships with.
Everything a lending team needs to turn a borrower's document pile into a moving, originated loan — plus a community to grow in.
# 01Any-document ingestion
PDFs, scans, phone photos, multi-doc packets — quality-agnostic, however messy the source.
ingest · split
# 02Auto-classification
Detects the document type and separates combined files into their parts automatically.
classify · route
# 03Field & table extraction
Pulls structured fields, line items, and tables — each with a confidence score.
fields · tables
# 04Validation & review
A rules engine plus human-in-the-loop review for anything below the confidence threshold.
rules · hitl
# 05Roles & multi-tenancy
Loan officers, processors, and tenant admins — isolated per lending organization.
rbac · tenant
# 06Social finance feed
Posts, comments, and likes — a community layer where finance pros share and discuss.
posts · social
# Engage
Let's talk growth.
30-minute scoping call. You leave with a written scope and a target ship date — or with an honest “we're not the right firm.”