Cuca-Insider-AI

dashboard demo sf8
chatboxt demo sf8
chatbox showing the process mode sf8
Load context analysis sf8 demo

Inspiration

While addressing Shinhan Future's Lab challenge "[SF8] AI-based Customer Behavior Prediction (New Customers)", we identified a major pain point: Telesales teams struggle to confidently pitch to "thin-file" customers who lack a traditional credit history. Rather than relying on static scoring formulas, we focused on actions, habits, and purchasing psychology. By analyzing behavioral statistical algorithms and seasonal cycles from 3rd-party digital footprints (Telco, Social, E-wallet), we aimed to build a tool that transforms raw data into actionable advisory scripts that hit the customer's true needs.

What it does

Cuca-Insider-AI operates as a smart lookup assistant for SVFC agents. When an agent searches for a new customer ID, the system processes their external behavioral data and displays it via intuitive visual reports. Instead of just showing numbers, the AI generates a customized advisory prompt. It explicitly recommends core SVFC products—specifically "Vay tín chấp cá nhân" (Personal Unsecured Loan) or "Thẻ tín dụng THE FIRST" (THE FIRST Credit Card)—based on lookalike customer segments.

For example, the AI will output: "12% of customers with a similar e-commerce spending pattern successfully opened THE FIRST credit card. Pitch this as an optimal gift-buying solution for the upcoming year-end holidays."

How we built it

We designed the software with a dual-flow information architecture, strictly focusing on "Harness Engineering":

Flow 1 (Deterministic): The software automatically calculates and outputs quantitative tables and charts for both cohorts and individual customers, based on a simulated 50-customer dataset mapped from alternative behaviors (E-wallet, Social).

Flow 2 (Generative AI): Driven by the statistical tables, the AI acts as an expert who has learned historical patterns to provide advice.

Example 1: Approaching the year-end shopping season, the system prompts the agent to pitch THE FIRST credit card to cohorts with high shopping affinity.

Example 2: If Customer A has maxed out their credit limit for 3 consecutive months but shows a solid declared income -> the AI suggests advising a card tier upgrade or a Personal Unsecured Loan.

Challenges we ran into

The core challenge was ensuring data privacy, preventing the AI from hallucinating fake statistics/products, and mitigating the risk of the AI altering or deleting system data. We resolved this through a strict 4-Layer Architecture:

Layer 1 (Native Calculation): Utilizing internal software algorithms to lock in absolutely accurate statistical numbers.

Layer 2 (PII Masking): Encrypting all Personally Identifiable Information (Name, Address, Bank Account) before model inference to ensure security and prevent AI context pollution.

Layer 3 (AI Qwen Inference): The Qwen model is strictly confined by rule-based prompts, forcing it to cross-reference behaviors with SVFC's official product catalog. Qwen is granted absolute "Read-only" access. Crucially, it quantifies "qualitative traits" (e.g., susceptible to flattery, impulsive spender, business owner) to warn of risks or identify upsell opportunities. The AI is also programmed to ask clarifying questions or alert the user if it detects anomalous input data.

Layer 4 (Reversion & Display): Decrypting the masked data and merging it with Layer 3's output to display on the agent's Dashboard/Chat UI.

Accomplishments that we're proud of

We successfully shifted the paradigm from a passive "Risk Information Dashboard" to a proactive "Sales Management & Enablement Tool." The system effectively bridges abstract 3rd-party data points (e.g., high telco data usage combined with continuous e-wallet top-ups) into a logical insight, translating directly into a Shinhan-specific product pitch.

What we learned

We learned that the true power of LLMs in fintech lies not only in predictive analytics but in Explainability. Providing telesales agents with the "Why" (the behavioral rationale and the 12% lookalike statistic) empowers them with confidence, smooths out call navigation, and ultimately drives a higher contract success rate.

What's next for Cuca-Insider-AI

Moving from the current Proof of Concept (PoC) stage, our technical Software Development Life Cycle (SDLC) roadmap includes:

Scaling the Data Pipeline: Transitioning from static mock data to building robust ETL (Extract, Transform, Load) pipelines for real-time behavioral data streaming.

Building a Feedback Loop: Implementing a feature for telesales agents to rate the AI's advice, utilizing RLHF (Reinforcement Learning from Human Feedback) so Qwen continuously adapts to real-world sales outcomes.

User Acceptance Testing (UAT): Packaging the MVP to run a pilot program with a select group of Shinhan agents, measuring the actual conversion Lift rate before a full-scale CRM integration.