ChihChi — AI Engineer

## 01 — about.py

Who is ChihChi?

chihchi.py
1class ChihChiKuo(AIEngineer):
2
3  title = "AI Engineer | Data Scientist Engineer"
4  company = "Cathay Life Insurance"
5  education = "NCTU · M.S. Industrial Engineering"
6  focus = ["Agentic AI", "LLM Evaluation", "MLOps"]
7
8  def build(self):
9    return [
10      "insurance product chatbot",
11      "AI voice agent MVP",
12      "internal LLM coding assistant",
13      "OCR / AIOps / anomaly detection"
14    ]
15
16  def mission(self):
17    return "Turn AI exploration into production service"

Agentic AI SystemPython LLM EvaluationDocker Voice AgentModel API OCR ModelServerless Anomaly DetectionCloud Native

Hi, I'm ChihChi Kuo — an AI Engineer and Data Scientist Engineer based in Taiwan, currently working in the financial industry on an AI & Data Team.

I specialize in moving enterprise AI ideas from early exploration into validated, deployable services. My work covers requirement clarification, technical selection, Agent application design, system design, software development, deployment, and evaluation.

Recently, I worked on an insurance product chatbot and improved the document retrieval stage to recall@15 = 0.98. I also helped design and build an AI Voice Agent MVP using OpenAI Realtime model concepts for inbound premium-related service calls.

Before focusing on LLM and Agent systems, I built OCR, time-series anomaly detection, AIOps, real-time and batch ML APIs, and production-grade model deployment workflows.

## experience

2025 – Now	Data Scientist Engineer Cathay Life Insurance · AI / LLM / Agent applications
2021 – 2025	Data Engineer Cathay Life Insurance · ML services, OCR, AIOps, internal LLM tooling
2019 – 2021	Data Scientist Taiwan Cement Corporation · anomaly detection and regression modeling
2017 – 2019	M.S. Industrial Engineering and Management National Chiao Tung University

## 02 — work/

Selected Work

★ FEATURED WORK → LinkedIn Profile

Insurance Product AI Assistant

Collaborated with a multi-developer team on a proof-of-concept AI assistant for insurance policy clauses. Built a knowledge-retrieval-driven Q&A workflow and optimized retrieval strategy to improve recall while keeping latency stable.

LLMRAGKnowledge Retrieval EvaluationInsuranceProduction AI

0.80→0.98

// recall@15 improvement

2026 Q1

// planned launch window

stable

// no latency increase

01

AI Voice Agent MVP Agent

Led development of a voice agent MVP for premium-related inbound calls using OpenAI Realtime model concepts. Designed dialogue flow, state-flow controls, and quality metrics for response quality, task completion, and interaction experience. Includes a live demo video below.

Realtime ModelVoice AgentState MachineLLM Evaluation

16.4%→8.6%

02

Internal LLM Coding Assistant LLM

Built an internal large language model service with a VS Code extension to provide GitHub Copilot-like assistance for IT teams while reducing risks around sensitive data and personal information leakage.

LLMVS Code ExtensionDeveloper ToolingPrivate AI

→ LinkedIn

03

OCR Key Information Extraction CV

Built and improved OCR models for key information extraction in business operations, raising online model accuracy from 81% to 93%.

OCRComputer VisionPythonModel Optimization

81%→93%

04

AIOps & Time-Series Anomaly Detection MLOps

Developed anomaly detection algorithms for life-insurance core system server metrics, cement plants, and battery factory scenarios, including algorithms that alerted operators 20 minutes earlier than existing rules of thumb.

AIOpsTime SeriesAnomaly DetectionPreventive Maintenance

20 min earlier

05

Production ML API Deployment Infra

Designed ML model systems including code structure, Docker containerization, deployment workflow, automated monitoring, and information visualization. Deployed 10+ real-time and batch ML models as APIs for internal systems.

DockerFlask APIK8sMonitoring

10+ models

## 03 — demos/voice-agent

AI Voice Agent Demo

// project

Premium-related Inbound Call Voice Agent

// architecture

OpenAI Realtime API · tool calling · state-machine flow · LLM-as-a-judge

// result

Unstable dialogue cases reduced from 16.4% to 8.6%

// description

A practical voice-agent MVP for premium-related service scenarios, redesigned from prompt-only behavior into a modular Voice Agent + Tool Calling system. The flow uses guardrails, conversation-state tools, intent identification, follow-up classification, and post-session scoring.

$ open ./voice_agent_demo

✓ Click to launch modal player

Platform: YouTube · Scenario: inbound service call