Cuca-Insider-AI

[SF8] Customer Behavior Prediction, Sales Closing Assistant

Comment

Inspiration

  • Telesales teams at Shinhan struggle to pitch confidently to "thin-file" new customers with no credit history. Traditional scoring models only show what might happen—but not how to approach the customer. We built Cuca-Insider-AI to transform raw alternative data (Telco, E-wallet, E-commerce, Social) into actionable advisory scripts that give agents the "why" behind every recommendation.

What it does

Cuca-Insider-AI is a smart lookup assistant for SVFC agents. When an agent searches for a customer, the system:

  1. Analyzes behavioral data from 3rd-party digital footprints
  2. Generates a tactical Next Best Offer (NBO) script via Qwen AI
  3. Outputs recommended_product, behavioral_rationale, statistical_evidence, sales_pitch_script, and risk_warning_and_upsell
  • For example: "12% of customers with similar e-commerce spending patterns successfully opened THE FIRST credit card. Pitch this as an optimal gift-buying solution for year-end holidays."

How we built it

A strict 4-layer privacy pipeline:

  • Layer 1 (Native Calculation): Deterministic stats and behavioral tags from alternative data
  • Layer 2 (PII Masking): Full name and phone removed before AI inference
  • Layer 3 (Qwen AI Inference): Rule-based prompts confined to SVFC product catalog (read-only)
  • Layer 4 (Reversion & Display): Merged output rendered in agent Dashboard/Copilot UI

Challenges

  • Preventing AI hallucination, ensuring data privacy, and keeping the AI read-only. Solved through strict prompt engineering and a 4-layer architectural boundary.

What we learned

  • The power of LLMs in fintech lies not in prediction alone, but in explainability. Giving agents the behavioral rationale and statistics empowers confident call navigation and higher conversion rates.

What's next

  • Scaling to real-time ETL pipelines, RLHF feedback loops from agent ratings, and UAT pilots with Shinhan agents before full CRM integration.
  • Fully integration on Cuca App Consumer: https://kts-04-cham-web-preview.vercel.app/ sh admin@test.com, Admin123!

Built With

Share this project:

Updates