Uploaded July 2026 | Updated September 2026, 2 weeks ago
LG Uplus handles massive customer service volume across South Korea: millions of subscribers, thousands of human consultants, and roughly 150,000 calls per day.
In this talk, MinkYu Ha from LG Uplus shares how the team built an AI Contact Center platform to support customers and human agents across the full service journey — before, during, and after a call.
MinkYu walks through the product architecture behind LG Uplus’s chatbot, callbot, real-time AI advisor, post-call summarization, CRM automation, and automated quality assurance. He also shares the technical lessons behind making these systems work in production, including document parsing for messy internal knowledge bases, layout-aware extraction, domain-specific embedding models, rerankers, hard negative training, agentic RAG, supervisor-agent patterns, and cost optimization with small language models.
The talk closes with practical lessons on evaluation-driven development, observability, feedback loops, and why production AI teams need to make quality measurable before they can improve it.
Chapters:
00:00 Intro and agenda
00:39 Why LG Uplus built an AI contact center
01:29 LG Uplus scale: subscribers, contact centers, and call volume
02:20 AICC results: self-service resolution and consulting time savings
03:05 The end-to-end AI customer service journey
03:47 Chatbots and callbots before the call
04:28 Real-time AI advisor during customer calls
05:17 Post-call summaries, CRM updates, and auto-QA
06:07 The knowledge base challenge: 30,000 changing documents
07:02 Why telecom documents are hard for AI systems
08:04 Why OCR and LLM cleanup were not enough
09:00 Building better document parsing for retrieval
09:52 Why benchmark scores do not guarantee domain performance
10:45 Custom embedding models and rerankers
11:35 Hard negatives and domain fine-tuning
12:25 Moving from advanced RAG to agentic RAG
13:24 Using memory, planning, and tools for better answers
14:15 Supervisor-agent pattern: reason, route, respond
15:08 Specialist agents for telecom knowledge domains
15:58 Reducing LLM cost at production scale
16:50 Why small language models work for agent tasks
17:42 Training small models with teacher trajectories and fine-tuning
18:44 Lessons learned from production AI
19:28 Evaluation-driven development and measurable quality
20:18 Observability, traces, feedback, and continuous improvement
Subscribe to Arize for more talks on AI agents, LLM evaluation, AI observability, agentic RAG, and production AI systems.
🔗 Try Arize AX & Phoenix OSS: arize.com
🔔 Subscribe for weekly content on LLMs, agents, and evaluation: youtube.com/@arizeai?sub_confirmation=1
LG Uplus handles massive customer service volume across South Korea: millions of subscribers, thousands of human consultants, and roughly 150,000 calls per day.
In this talk, MinkYu Ha from LG Uplus shares how the team built an AI Contact Center platform to support customers and human agents across the full service journey — before, during, and after a call.
MinkYu walks through the product architecture behind LG Uplus’s chatbot, callbot, real-time AI advisor, post-call summarization, CRM automation, and automated quality assurance. He also shares the technical lessons behind making these systems work in production, including document parsing for messy internal knowledge bases, layout-aware extraction, domain-specific embedding models, rerankers, hard negative training, agentic RAG, supervisor-agent patterns, and cost optimization with small language models.
The talk closes with practical lessons on evaluation-driven development, observability, feedback loops, and why production AI teams need to make quality measurable before they can improve it.
Chapters:
00:00 Intro and agenda
00:39 Why LG Uplus built an AI contact center
01:29 LG Uplus scale: subscribers, contact centers, and call volume
02:20 AICC results: self-service resolution and consulting time savings
03:05 The end-to-end AI customer service journey
03:47 Chatbots and callbots before the call
04:28 Real-time AI advisor during customer calls
05:17 Post-call summaries, CRM updates, and auto-QA
06:07 The knowledge base challenge: 30,000 changing documents
07:02 Why telecom documents are hard for AI systems
08:04 Why OCR and LLM cleanup were not enough
09:00 Building better document parsing for retrieval
09:52 Why benchmark scores do not guarantee domain performance
10:45 Custom embedding models and rerankers
11:35 Hard negatives and domain fine-tuning
12:25 Moving from advanced RAG to agentic RAG
13:24 Using memory, planning, and tools for better answers
14:15 Supervisor-agent pattern: reason, route, respond
15:08 Specialist agents for telecom knowledge domains
15:58 Reducing LLM cost at production scale
16:50 Why small language models work for agent tasks
17:42 Training small models with teacher trajectories and fine-tuning
18:44 Lessons learned from production AI
19:28 Evaluation-driven development and measurable quality
20:18 Observability, traces, feedback, and continuous improvement
Subscribe to Arize for more talks on AI agents, LLM evaluation, AI observability, agentic RAG, and production AI systems.
🔗 Try Arize AX & Phoenix OSS: arize.com
🔔 Subscribe for weekly content on LLMs, agents, and evaluation: youtube.com/@arizeai?sub_confirmation=1
