• 

  Lending analytics

• 

  SME rating gauge

• 

  Bill management

• 

  Employee management

• 

  Home page

• 

  Loan

• 

  Order management

• 

  Warehouse management

• 

  human's cleverness wins limitations

• 

  Vibe coding? I vibe engineering, took 10 years to fix, which is fine, job secured!

Shinhan Soft POS

The Merchant Operating System with Embedded Credit

Shinhan Future's Lab Hackathon

---

The Problem

Vietnamese SME Merchants Struggle Every Day

Daily Pain Points	Current Reality
Managing POS, Shopee, TikTok, MoMo separately	3 different apps for sales channels
Counting inventory manually every evening	Excel spreadsheets for inventory
Tracking staff hours and shift schedules	Paper timesheets for staff
Reconciling payments from 5+ platforms	2+ hours nightly reconciliation
Cash flow gaps between sales and settlements	No visibility when cash will run out

---

The Credit Gap

What Banks Want	What SMEs Have
6 months statements	Cash registers (POS)
Tax returns	Shopee/Lazada/TikTok sales
Collateral	MoMo wallet transactions
Formal business registration	Daily sales receipts

40.5% of Vietnam's firms have bank loans - World Bank Enterprise Survey

---

The Solution

Shinhan Soft POS

The Alternative Credit Layer for Vietnam's SMEs

Alternative Credit Foundation

• POS + marketplace + e-wallet data fusion
• ML-based normalization turns fragmented activity into one merchant timeline
• Alternative credit profile built from explainable transaction patterns
• Signals banks can actually underwrite against

Embedded Capital Experience

• Embedded working capital at the point of need
• 72-hour approval target with policy-based pre-screening
• No traditional bureau file required
• Human-reviewed underwriting inside daily utility

---

Daily Utility

This is what merchants do EVERY DAY

🌅 Morning	☀️ Midday	🌆 Evening	📈 Any Time
Check overnight orders; review inventory levels; assign staff shifts	Process in-store sales; fulfill marketplace orders; print customer receipts	Auto-reconcile all payments; see low-stock alerts; close shift reports	Check sales performance; manage staff; restock supplies

---

Engagement Thesis

Why merchants come back every single day

❌ This is NOT a loan app

Merchants don't install loan apps. They install tools that run their business.

✅ This is their operating system

• Their business lives inside this app
• Trust is built through daily usage
• Transaction data accumulates naturally
• Financing appears only when they need it

---

Embedded Capital

Financing when it matters, not as an ad

Contextual Triggers

• 📦 Low stock alert -> Restock financing
• ⏳ Settlement delayed -> Cash flow advance
• 📈 Sales growing fast -> Growth capital

How it works

• Based on real business performance
• No separate application
• No credit check required
• Fast, transparent, fair terms

---

The Loan Product

Clean loan terms for eligible merchants

Loan Range	$200 - $5,000
Interest Rate	15-18% APR
Term Length	3, 6, or 12 months
Approval Time	72 hours
Min. History	90 days

---

Demo Flow

How merchant activity turns into financing

1. Step 1 - Install SOL POS -> Merchant starts using the operating tool every day.
2. Step 2 - Data Normalization -> Transactions and consented mobile or telco signals are cleaned and standardized.
3. Step 3 - Feature Profile Forms -> Audit-ready transaction features populate the scorecard profile.
4. Step 4 - 90 Days of History -> Seasonality, stability, and risk features become reliable.
5. Step 5 - Policy Eligibility -> Policy rules + score threshold determine eligibility for underwriter review.
6. Step 6 - Loan Offer Shown -> Illustrative amount and terms appear inside the workflow.
7. Step 7 - Funds in 72 Hours -> Approved merchants receive funds after underwriting and policy checks.

---

AI Governance Framework

Explainable scoring with human review by default

Human-in-the-Loop as Default

• Every score/recommendation -> human review
• Fixed reason codes accompany each output
• LLM translates only approved reason codes
• Inputs, weights, bins logged for audit

What AI Actually Does (scoring path)

• Transaction cleaning and categorization
• Merchant behavior feature engineering
• WoE logistic scorecard -> primary model
• EBM challenger -> validation only

What AI Does Not Do

• Final credit decisions (bankers decide)
• Custom explanations (approved codes only)
• Black-box inference in scoring path

---

Explainable Credit Intelligence Layer

Where Shinhan's underwriting logic stays transparent

Capability	What it does
ML-Based Data Normalization	POS, marketplace, e-wallet, and telco metadata are fused into one audit-ready merchant timeline.
WoE Scorecard Core	Fixed bins, weights, and reason codes power the primary scorecard. An EBM challenger runs in parallel for accuracy checks.
Statistical Pattern Features	Tet, monsoon, and household-business effects become peer-normalized features inside the scorecard.
AI-Assisted Underwriting	Policy rules and score thresholds create a recommendation pack. Human review stays standard, and the LLM only translates approved codes.

Every score is reproducible from logged inputs, weights, bin boundaries, and reason codes.

---

The Data Flywheel

The Data Flywheel

More merchants -> More data fused -> Better scorecards -> Faster underwriting -> Better risk features -> Lower defaults -> Better portfolio -> Lower cost of capital -> Lower rates -> More demand

Performance data improves the next validation cycle.

---

Competitive Advantage

	Traditional POS	Marketplace Apps	Bank Loans	Shinhan Soft POS
Multi-channel orders	❌	✅ Only own channel	❌	✅ All channels unified
Inventory management	⚠️ Basic	❌	❌	✅ Smart alerts
Staff management	❌	❌	❌	✅
Automatic reconciliation	❌	❌	❌	✅
Contextual financing	❌	❌	❌	✅ Embedded
Daily usage frequency	Daily	Weekly	Monthly	Multiple times per day

Sources: MISA Lending (misa.vn), Shopee SEasy (TPBank partnership), Validus Vietnam, Techcombank ShopCash

Why This Matters

• Existing solutions use one data channel only
• Our approach combines POS + Marketplace + E-wallet data - not yet disclosed as a combined approach in Vietnam

---

Payment Safety

How we protect merchants and Shinhan

3 statistical safety patterns we monitor:

🔄 Circular Wallet Flows	🛒 Synthetic Orders	⚡ Velocity Anomalies
Rule-based flow checks flag repeated wallet loops that appear to inflate transaction volume.	Cross-checks compare payment, address, and fulfillment trails to surface synthetic orders.	Threshold and peer-baseline checks surface unusual spikes at odd hours or from new devices.

---

Vietnam-Specific Risk Signals

Peer-normalized statistical signals tuned to how local merchants actually operate:

🎊 Tet Season Patterns	🌧️ Monsoon Disruption	👨‍👩‍👧 Family Business Structures
Lunar-calendar seasonality factors help prevent holiday spikes from being mistaken for permanent growth.	Regional baseline adjustments separate weather-driven demand drops from true merchant deterioration.	Linked-counterparty features capture shared household cashflow patterns while keeping the score fully auditable.

---

Compliance Roadmap

Path to Production

Phase 1	Phase 2	Phase 3
Sandbox (Month 1-2): Anonymized data only; internal testing; score + reason codes prepared for analyst review; SBV regulatory consultation	Pilot (Month 3-4): Live data with full consent; POS operations first; automated score calculation and policy checks; human review remains standard; compliance audit trail	Production (Month 5-6): Full POS rollout; contextual capital offers; explainable scorecard + audit trail; human-reviewed origination at 100+ merchants

---

The Ask

What we need

• Access to merchant/payment infrastructure
• Partnership with existing POS providers
• Pilot merchant cohort for validation
• 6 months to prove daily engagement

What you get

• The daily merchant operating system for Vietnam
• High frequency merchant relationship
• Embedded financing monetization layer
• Data asset that compounds with every transaction

---

What This Is

We haven't launched a production credit product yet.

We don't have live API integrations yet.

We don't have a backtested model yet.

What we HAVE built:

• Working early system showing normalized data -> explainable score concept
• ML-based data fusion and scorecard pipeline architecture
• Clear path to 6-month early product launch

This is an early product proposal with a defined roadmap.

---

The Close

Every sale is captured.

Every inventory alert is acted on.

Every merchant grows with Shinhan.

Position Shinhan as the bank merchants use every day.

---

Appendix: Research Sources

Market Context

Source	What It Verifies
Vietnam News - Digital Loans Open New Doors for SMEs	SMEs struggle to access bank credit; digital lending growing
Techcombank - MISA Lending	Existing bank alternative-data lending product
MISA Lending - MB Partnership	Digital SME lending via accounting software

Existing Competitor Products

Source	What It Verifies
Shopee Seller Education - SEasy	Shopee seller financing product exists
Validus Vietnam	Active SME lender in Vietnam
Techcombank MerchantOne	Bank merchant acceptance product
Techcombank SoftPOS	Bank SoftPOS product
Techcombank ShopCash	Bank merchant lending product

---

Appendix: Model Explainability

WoE Scorecard & EBM Challenger

WoE Logistic Scorecard

Converts transaction history into a single score (300-850 scale) with fixed reason codes. Every input, weight, and bin boundary is logged and reproducible. Regulators prefer it because every decision is fully auditable.

EBM Challenger Model

A glass-box ML model that runs alongside the scorecard for accuracy validation. Each feature's contribution to the final score is visible. Used for validation only - not final decisions.

---

Why Not Deep Learning or LLM for Scoring?

Black-box models are prohibited by most banking regulators - no one can explain why a merchant was declined. Our LLM is limited to translating pre-approved reason codes into natural language. It never makes or modifies credit decisions.

Every score is reproducible from logged inputs, weights, bin boundaries, and reason codes.

Built With

• adenosin-triphosphaste
• alibaba-model-studio
• bmad
• claude-sonnet-4.6
• cloudflare-workers
• coffee
• dola-seed-2.0-pro
• gpt-5.4
• grok-code-fast-1
• groq
• groq-compound-mini
• kilo
• magic
• minimax-m2.5-highspeed
• minimax-m2.7-highspeed
• neovim
• openrouter
• puter-js
• qwen-3.6-plus
• qwen3-coder
• qwen3-coder-480b-a35b-instruct
• react-router
• tailscale
• termux
• vibe-engineering
• waifu
• zellij