User side
Bank Admin side
Tech Stack
Our Data

Lumina Bank

User: https://lumina.trinova.it.com

Bank Admin: https://lumina.trinova.it.com/admin_dashboard

[deployed on Alibaba ECS then took down to save maintenance cost]

Inspiration

Shinhan signal	Why it mattered for this project
Global network in 20 countries with major presence in markets including South Korea, Vietnam, Japan, the U.S., and Indonesia	Suggested a real need for smoother cross-border document flow, stronger shared case visibility, and faster coordination across customer-facing and internal systems at global scale.
Global operations	Reinforced the need for structured, repeatable, document-heavy workflows that cannot rely only on manual review

That led us to one core question:

What if customer-submitted loan documents could move from upload to OCR to underwriting support to admin review in one clean AI-assisted flow?

What it does

Lumina Bank is a multi-agent AI copilot for autonomous document processing and underwriting workflows.

Designed as a financial-services AI product for Lumina Bank-style loan and credit operations, it acts as a reviewer-ready, human-in-the-loop copilot that helps teams intake documents, extract evidence, surface risk flags, route cases, and prepare cleaner underwriting decisions.

Pain points we targeted

Pain point	How Lumina Bank helps
Manual document collection	Centralizes intake into one structured workflow
Fragmented evidence checks	Pulls OCR, evidence, and risk signals into one review experience
Repetitive follow-ups	Supports clarification and re-upload loops
Hand-built case summaries	Generates reviewer-ready summaries and reports
Disconnected customer/admin workflows	Links customer uploads directly to admin review context

What Lumina Bank supports

Capability area	Support provided
Document intake	Customer upload and submission flow
OCR and structured extraction	Raw files converted into usable fields and evidence
Evidence retrieval	Search and retrieval across submitted content
Fraud and inconsistency checks	Flags mismatches and suspicious patterns
AI-assisted loan insights	Recommendation support and next-best actions
Admin visibility	Customer-uploaded documents visible in review flow
Routing and review support	Queueing, escalation, and triage support
Human-in-the-loop approval	Final decision remains with the reviewer

How we built it

We built Lumina Bank as a full-stack workflow prototype spanning both the customer portal and the admin portal.

Core stack

Layer	Technology
Frontend	React, TypeScript
Backend	FastAPI, Python
Storage	Alibaba Cloud OSS
Persistence	Backend-managed document metadata and application linkage
AI / OCR	Qwen-based OCR and document analysis
Fallback model	OpenAI GPT-4.1 mini

AI pipeline and workflow structure

We designed the system as a staged workflow for:

document upload and intake
OCR extraction and normalization
evidence retrieval
fraud detection
loan recommendation
narrative generation
admin review and decisioning

Security and architecture choices

Choice	Why it matters
Backend-mediated uploads	Browser never receives permanent Alibaba credentials
Private OSS storage	Uploaded documents stay protected by default
Backend-managed metadata	Documents stay linked to applications and workflow state
Signed temporary URLs	Secure access without exposing the bucket publicly

That let us build a flow that is both demo-ready and production-shaped.

OpenClaw and NemoClaw work

Area	What we did
OpenClaw product layer	Implemented an OpenClaw orchestration experience inside the Support tab, where one supervisor flow coordinates case context, evidence boundaries, customer follow-up drafts, and admin-facing summaries
NemoClaw / OpenShell runtime work	Performed real local bring-up work including Docker validation, gateway recovery, onboarding progress, and sandbox creation attempts
Outcome	Full local sandbox creation on macOS was blocked by a Docker Desktop/OpenShell build-stream incompatibility, but the work still informed the architecture and helped us distinguish app-level orchestration from sandboxed runtime paths

Multi-agent architecture

Agent	Responsibility
Intake agent	Document collection, quality checks, and intake completion
Extraction agent	OCR, normalization, and structured field generation
Evidence agent	Semantic search and source-grounded retrieval
Fraud agent	Inconsistency checks and explainable risk flagging
Insight agent	Underwriting support, recommendation logic, and next-best actions
Report agent	Case-brief generation and reviewer summaries
Routing agent	Queue assignment, escalation, and triage decisions
Notification agent	Clarification requests, follow-ups, and communication logging

This architecture lets the product move beyond simple OCR and become a full underwriting workflow assistant.

Core features

Feature	Description
Dual document ingestion	Accepts image and PDF uploads, with room for clarification or re-upload when documents are incomplete or unreadable
Intelligent OCR & structured extraction	Converts raw files into normalized, reviewer-friendly fields and extracted evidence
Semantic search / evidence retrieval	Helps reviewers quickly find supporting details across a document set
Fraud and inconsistency detection	Surfaces explainable mismatches, missing evidence, suspicious patterns, and review-triggering anomalies
AI loan insights & recommendation	Produces suggested loan range, rationale, confidence cues, and next-best action guidance
Auto report generation	Packages findings into a reviewer-ready case brief
Autonomous workflow routing	Directs cases into approval, escalation, clarification, or manual review queues
Human-in-the-loop review	Keeps final judgment with reviewers and admins

System workflow

Customer-side flow

Step	Description
1	Applicant submits supporting documents through upload or live capture
2	System validates file presence and quality
3	OCR and structured extraction convert documents into usable fields and evidence
4	If information is missing or unclear, the system can trigger a clarification or re-upload loop
5	A case package is assembled for underwriting review

Reviewer / admin-side flow

Step	Description
1	Underwriter or admin receives a triaged case in the dashboard or work queue
2	Retrieved evidence, extracted fields, and explainable risk flags are reviewed together
3	System proposes loan insights, recommended ranges, and next-best actions
4	A reviewer-ready brief is generated for faster decisioning
5	Human reviewers approve, reject, escalate, or request more information while maintaining a communication trail

Accomplishments that we're proud of

We are proud that this became much more than a static UI concept.

Workflow accomplishments

Accomplishment
Built a real Loans page → backend → Alibaba OSS upload path
Linked uploaded customer documents to application context
Made customer-uploaded documents visible in the admin review flow
Created a true end-to-end journey from application submission to admin decision
Supported customer status tracking after submission

AI and review accomplishments

Accomplishment
Built a workflow covering OCR, fraud checks, loan recommendation, and narrative generation
Exposed OCR output in reviewable formats instead of hiding it behind a black box
Added fallback behavior so the system is more resilient in demos and testing
Kept the final decision with the human reviewer while still providing AI guidance

Infrastructure accomplishments

Accomplishment
Preserved private OSS bucket access
Implemented signed-URL retrieval instead of public file links
Exposed teammate-friendly metadata and retrieval endpoints
Kept the architecture practical enough for future extension

Orchestration accomplishments

Accomplishment
Built an OpenClaw orchestration flow in the Support tab
Connected supervisor-style outputs such as decision summaries, customer notification drafts, and admin case reports
Investigated real NemoClaw/OpenShell runtime onboarding instead of only describing it conceptually

We are especially proud that Lumina Bank lands in a realistic middle ground:

not replacing the underwriter, but helping the underwriter work faster with better-organized inputs and stronger decision support.

Built for the following requirements

Requirement area	How Lumina Bank addresses it
High-accuracy OCR & data extraction	OCR workflow and structured extraction pipeline
Loan analysis process automation	Multi-stage AI pipeline
Fraud detection	Biometric and rule-based checks
Loan insights recommendation	Narrative and recommendation layer
Loan amount recommendation	Range recommendation logic
Admin decision support	Review-ready summaries and evidence context
Customer-to-admin workflow continuity	Linked uploads, visibility, and case flow

What's next for Lumina Bank

Near-term next steps

Priority area	Next step
OCR quality	Richer structured extraction for financial forms and tables
Normalization	Better post-processing for dates, IDs, amounts, and rates
Reliability	Stronger multi-model review for low-confidence fields
Admin support	Deeper summarization and recommendation support
Reporting	Improved automated report generation for underwriting handoff
Auditability	Clearer tracking around extraction, fallback, and review behavior

Longer-term vision

Vision area	Goal
Underwriting workspace	Stronger AI-assisted review environment
Deployment	More robust production patterns for storage, preview, and access control
Customer workflow	Better document re-upload and clarification loops
Scale	Review tooling across multiple financial products and internal teams
Orchestration/runtime	Continued OpenClaw/NemoClaw evolution in a more reliable sandbox/runtime environment

Longer term, we see Lumina Bank as an AI-assisted loan operations platform that helps Shinhan-style financial workflows move faster while still keeping humans in control.

Tech stack

Category	Stack
Frontend	React, TypeScript
Backend	FastAPI, Python
Storage	Alibaba Cloud OSS
AI / OCR	Qwen-based document analysis workflows
Fallback model path	OpenAI GPT-4.1 mini
Orchestration layer	OpenClaw-inspired supervisor flow
Runtime investigation	NemoClaw / OpenShell bring-up and sandbox research
Database / metadata	Backend-linked application and document persistence