iThome Hello World 2025
At this year’s iThome Hello World 2025 Dev Conference, I delivered four back-to-back sessions focused entirely on the practical realities of putting AI into production. This wasn’t just about sharing knowledge; it was a distillation of the hard-learned lessons our team gathered while deploying LLMs in the banking sector.
Rather than pitching grand visions of the AI future, these four sessions zeroed in on how to build AI systems that are stable, deep, and secure. Here’s a quick recap of the sessions:
1. The MCP Major League: Building AI-Driven Data Integration and Early Prompting
(Oct 14, 2025 | 11:00 - 12:30)
The Model Context Protocol (MCP) is redefining how AI Agents connect with external systems. In this session, we explored how to leverage MCP to build a standardized toolchain, allowing LLMs to securely and efficiently call internal enterprise data APIs, while using Early Prompting mechanisms to guide the agent’s decision pathways.
2. Making Memories “Flock Together”: Using GraphRAG for Customized AI Customer Service
(Oct 14, 2025 | 15:25 - 15:55)
Traditional RAG often falls short when faced with cross-dimensional, multi-hop reasoning questions. We shared our hands-on experience of upgrading from flat-text RAG to a knowledge graph architecture (Ontology + GraphRAG). By letting data structure itself into relational clusters, the customized AI can answer not just isolated queries, but demonstrate contextual memory and deep understanding.
3. Five Interns, One Assembly Line: Vibe Coding Practical Workshop
(Oct 15, 2025 | 09:00 - 10:30)
How do you lead newcomers with zero experience to quickly build functional AI prototypes? In this 90-minute hands-on workshop, I walked the audience through the agile “Vibe Coding” paradigm. We demonstrated how to use tools like Cursor, Claude, or Gemini to seamlessly bridge prompting and coding, elevating the productivity of five people to that of a full development squad.
4. Enterprise-Grade LLM Guardrails and Prompt Hardening
(Oct 15, 2025 | 14:10 - 14:55)
In the financial industry, security and compliance are non-negotiable. This session tackled the hardest battles in LLM defense — from preventing Prompt Injections to implementing enterprise-grade Guardrails. We explored how to use Prompt Hardening to strengthen the model’s inherent resistance, and how to build external validation layers to ensure stable, compliant AI outputs.
If you’re interested in the slides or the technical deep-dives from these sessions, feel free to check out the official speaker page or reach out for a discussion!
👉 View Full Agenda and Speaker Details (iThome Hello World 2025)
💬 Read more: 2025 Year in Review (English)
RAG System Design
Building a RAG system in banking: how to choose your chunk strategy, embedding model, and retrieval pipeline. Lessons from real production experience.
Why Deploying LLMs in Banking Is 10x Harder Than You Think
Deploying AI in a bank isn't just picking a model. Compliance, security, data governance, organizational culture — each hurdle is the necessary path from 1...
RAG System Design
Building a RAG system in banking: how to choose your chunk strategy, embedding model, and retrieval pipeline. Lessons from real production experience.
Why Deploying LLMs in Banking Is 10x Harder Than You Think
Deploying AI in a bank isn't just picking a model. Compliance, security, data governance, organizational culture — each hurdle is the necessary path from 1...