Ask the Intern

AI-powered reporting assistant for Shinhan that auto-tags banking transactions and lets any team query business data in natural language -- with human-in-the-loop review for compliance.

Comment

Inspiration

Every bank has the same bottleneck: business teams need data to make decisions, but they're stuck waiting days for a small reporting team to manually pull numbers from internal systems. At Shinhan Bank Vietnam, only the ICT Reporting Team has access to Power BI, and it's not widely adopted. Everyone else submits a request and waits.

We asked ourselves: what if any department head could just ask for the data they need, in plain Vietnamese or English, and get an accurate, compliant report back in minutes instead of days?

But we realized the reporting problem is actually a data quality problem in disguise. If transaction data isn't consistently tagged and categorized as it flows through the system, no amount of clever querying will produce reliable reports. So we decided to tackle both sides: clean data going in, and easy access coming out.

What it does

Ask the Intern is a two-layer AI system for Shinhan's internal teams:

Layer 1 - Smart Transaction Tagging: As banking transactions flow through the system, Qwen automatically classifies them by type, department relevance, tax category, and risk flags. This creates a clean, structured data foundation that doesn't exist today.

Layer 2 - Conversational BI: Authorized users ask business questions in natural language. The system translates these into structured database queries, returns results as visualizations and summaries, and routes everything through a human-in-the-loop review before distribution -- ensuring accuracy and regulatory compliance.

How we built it

  • Qwen 3.6 Plus via Alibaba Cloud Model Studio API for natural language understanding, transaction classification, text-to-SQL conversion, and report summarization
  • Qwen's multilingual support for Vietnamese and English query handling
  • Simulated banking dataset to demonstrate the full pipeline
  • Human-in-the-loop approval workflow for compliance

Challenges we faced

  • Designing a realistic transaction tagging taxonomy without access to Shinhan's actual internal data
  • Balancing query flexibility (natural language is messy) with output accuracy (banking reports must be precise)
  • Building a compliance review step that adds trust without adding friction

What we learned

  • The gap between "AI can answer questions" and "AI can answer questions a bank will trust" is enormous -- the human-in-the-loop design is not a nice-to-have, it's the whole point
  • Transaction auto-tagging is the unsexy foundation that makes everything else possible

What's next

  • Integration with Shinhan's actual data systems and reporting workflows
  • Role-based access controls aligned with Shinhan's internal security policies
  • Expanding to proactive alerts: the system notices trends and flags them before anyone asks

Built With

  • alibaba-cloud-model-studio-api
  • python
  • qwen-3.6-plus
  • sql
  • streamlit
Share this project:

Updates