From Enterprise AI Chatbots
to LLM-Powered Agent Systems

Why This Experience Maps to Wonderful

Direct alignment between Wonderful's FDE & Field CTO requirements and my 10+ year track record building and deploying AI systems for enterprise customers.

Enterprise Clients & Domain

Currently serving as CTO at Pevo across two enterprise engagements in South Korea's healthcare ecosystem, owning system architecture, database design, DevOps, MLOps, and security.

Enterprise AI in Production — Before Pevo

Left: Amore Pacific (Etude House) — "Color Picking Chatbot" on Facebook Messenger. AI-powered lip color recommendation via image recognition, deployed to 10,000+ users. Right: Gangnam District Office — Public service chatbot handling citizen inquiries via KakaoTalk.

IGC 2016 Conference — Presenting "AI Platforms for Game Customer Service". Demonstrated text-to-vector transformation for automated customer inquiry classification using CNN (Convolutional Neural Networks).

Three AI Agents, One Healthcare Platform

Each agent addresses a distinct operational challenge — from document processing to real-time voice interaction — connected through shared OCR/LLM infrastructure.

Agent 01 — Insurance

Automated Damage Assessment

Multi-agent system that automates the labor-intensive process of reviewing veterinary medical receipts and determining insurance coverage eligibility. Built for KB Insurance's pet insurance division, reducing claim processing time while maintaining accuracy through human-in-the-loop verification.

Data Extraction Pipeline — Vision AI + Medical AI agents digitize handwritten receipts into structured data, then RAG agents cross-reference against insurance contracts for coverage eligibility

Data Extraction Pipeline

OCR Module → LLM structured extraction. Handles receipts, ID cards, pet registration docs, and bank account documents. Vision AI + Medical AI agents digitize messy handwritten receipts into normalized JSON.

Coverage Analysis Engine

Strategy pattern per insurer (KB, DB, Meritz, Samsung, Hyundai). Common classification agent (1 LLM call) → per-insurer rule calculation (0 LLM calls). 9-category exclusion framework with surgical procedure linking.

Human-in-the-Loop Interface — Certified loss adjusters verify AI classifications via drag-and-drop, with automated data logging creating a continuous feedback loop for model improvement

Human-in-the-Loop

Certified loss adjusters review AI classifications via drag-and-drop interface. Override capability with structured reason codes. Every manual decision is logged to create a continuous feedback loop for model improvement.

Multi-Insurer Optimization

Reduced LLM calls from N (one per insurer) to 1 via common classification layer. Individual strategy calculations are pure rule-based — no LLM, no latency. DB-driven rules allow A/B testing without code changes.

Receipt Image → OCR Module → Text Lines → LLM Extraction → Structured Items
                                                              ↓
                                               Common Classification Agent  ← 1 LLM call
                                                    ↓              ↓
                                          covered_items    excluded_items
                                           ↓         ↓         ↓         ↓
                                      KB Strategy  DB Strategy  Meritz  Samsung  ← 0 LLM calls
                                           ↓
                                   Loss Adjuster Review → Approve / Reject / Override
                                           ↓
                                   Feedback Loop → Model Retraining Data

Agent 02 — Pharmacy

Automated Diabetes Prescription Processing

Extracts structured data from 20+ hospital-specific diabetes prescription formats using the same OCR + LLM pipeline as Agent 01 — then automatically submits reports to South Korea's National Health Insurance Service (NHIS). This agent's prescription analysis output directly powers Agent 03's personalized voice interactions.

Prescription Processing Pipeline — Vision AI extracts 30+ fields from 20+ hospital-specific diabetes prescription formats, with automated reporting to the National Health Insurance Service

Multi-Format Extraction

Auto-detects 3 prescription types (general diabetes, medical aid, CGM continuous monitoring). Type-specific LLM prompts extract 30+ fields: patient info, diagnosis codes (E10-E14), medication items, dosage schedules, insulin usage, and institutional codes.

NHIS Reporting Automation

Extracted patient info + diagnosis codes are automatically formatted and submitted to the National Health Insurance Service. Dynamic protocol adaptation based on diabetes type and level. End-to-end automation from paper prescription to government submission.

Prescription Image → OCR Module → Text → Type Auto-Detection
                                                    ↓
                               ┌────────────────────┼────────────────────┐
                         General Diabetes    Medical Aid       CGM Electrode
                               └────────────────────┼────────────────────┘
                                                    ↓
                                         LLM (type-specific prompt)
                                                    ↓
                                         Structured Prescription JSON
                                           ↓                    ↓
                                  NHIS Auto-Report     → Agent 03 (Voice AI)
                                                         medication data feeds
                                                         personalized calls

Agent 03 — Voice AI

Medication Adherence Voice Agent

A real-time voice AI agent that calls patients at scheduled times to verify medication intake, check for side effects, and escalate safety concerns to their prescribing pharmacy or hospital. Built on Agent 02's prescription analysis — the system knows exactly which medications each patient takes, their dosage schedule, and relevant drug interactions.

Medication Schedule — Time-based dosage tracking extracted from prescriptions via Agent 02

AI Call Settings — User configures call times and medication groups

Key Architecture Decision: Dual-Channel Design

Instead of streaming raw audio to the server (which degrades quality over mobile networks), the system performs on-device STT and sends only text over a dedicated text channel. A separate audio channel handles TTS playback from the server. This dual-channel WebSocket design minimizes latency and preserves voice quality — critical for elderly patients who make up the majority of medication adherence users.

VoIP Call Infrastructure

Server triggers FCM (Android) / APNs VoIP Push (iOS) at scheduled times. CallKit integration presents native phone UI — patients answer a "real" phone call. Cross-platform: iOS uses AVAudioEngine + CallKit; Android uses WebRTC + FCM data messages.

On-Device STT (Multi-Provider)

OpenAI Whisper, ElevenLabs Scribe, Google Cloud, Naver Clova, and iOS/Android native speech recognition. VAD (Voice Activity Detection) with RMS threshold tuning for speaker vs. earpiece modes. 0.7s silence detection triggers transcription; 5s max utterance limit.

LLM Conversation Engine

LLM with dynamic system prompt injected with patient's prescription data from Agent 02. Korean AI pharmacist persona with sliding window memory (last 10 turns). Structured conversation flow: greeting → medication check → side-effect screening → safety escalation → farewell.

Real-Time TTS Streaming

ElevenLabs (primary), Google Wavenet, Naver Clova as providers. Server generates TTS audio chunks → streams over WebSocket → native audio engine playback. Echo cancellation: microphone automatically muted during TTS playback, buffer cleared on TTS end. TTS caching strategy: pre-generates and caches TTS responses based on patient profile data (medications, schedule, common dialogue patterns), maximizing cache hits to deliver minimal latency during live calls.

End-to-End Call Flow

1. Scheduled Trigger

Server checks dosage schedule → sends VoIP push notification at configured time (e.g., 1:00 PM for morning + lunch meds)

2. CallKit / Native Ring

Patient sees incoming call from "AI Pharmacist" — standard phone UI, works on lock screen. Patient taps Accept.

3. Dual WebSocket Channels Established

Text channel: carries STT transcripts + LLM responses. Audio channel: carries TTS audio chunks. Both over WSS with session management.

4. Dynamic Conversation

AI: "Good afternoon! Did you take your Metformin 500mg with lunch today?" — generated from Agent 02's prescription data. Patient responds naturally; on-device STT converts to text.

5. Side-Effect Screening

LLM probes for common side effects based on prescribed drugs. "Have you experienced any nausea or dizziness since starting this medication?"

6. Safety Escalation

If risk signal detected (adverse reaction, missed doses, concerning symptoms) → system connects patient to prescribing pharmacy or hospital. Transcript saved for clinician review.

Dosage Scheduler → FCM / APNs VoIP Push → CallKit (iOS) / FCM (Android)
                                                        ↓
                                               Patient accepts call
                                                        ↓
                                        ┌───── Dual WebSocket ─────┐
                                        │                            │
                                  Text Channel              Audio Channel
                                        │                            │
                            On-device STT → text           TTS audio chunks
                                        ↓                            ↑
                              LLM + Prescription DB   →   ElevenLabs TTS
                              (dynamic prompt injection)
                                        ↓
                              Adherence logged to DB
                              Risk? → Pharmacy / Hospital alert

Full-Stack Architecture Ownership

Designed and operated the complete technical stack across both projects — from infrastructure and databases to ML pipelines and mobile client.

Production Results

Why Wonderful

What draws me to Wonderful is the local-first strategy. AI agents don't succeed through one-size-fits-all deployment — they succeed through deep local adoption, understanding each market's regulations, language nuances, and customer workflows. This is exactly how I've operated: navigating Korean healthcare regulations, building Korean-language voice AI, and adapting to local enterprise procurement processes. Wonderful's approach of embedding with customers locally to drive real adoption, rather than selling from a distance, resonates deeply with how I've built my career.

I'm especially excited about Wonderful's focus on Voice AI agents. Voice is rapidly becoming the defining trend in enterprise AI — and South Korea is the perfect proving ground. The Korean market is saturated with voice-based customer service operations across telecom, banking, insurance, and healthcare. These are high-volume, high-cost call centers ripe for AI transformation. Having already built a production Voice AI agent that makes VoIP calls, handles natural conversation, and integrates with backend systems, I know firsthand both the technical challenges and the massive business opportunity. Voice AI agents aren't just a feature — they're a standalone business model that can replace a significant portion of manual CS costs.

That's why I'm drawn to Wonderful: it's the work I've already been doing, the market I understand deeply, and the future I want to help build at global scale.