AI/Voice Engineer / Team Lead @ Soundmind AI Engineer @ WIGTN Crew

Building voice AI agents for real phone call environments. Leading the full cycle — STT/TTS model serving & optimization, dialogue state machine design, and model evaluation at scale.

I believe the gap between paper benchmarks and production performance is real, and the only way to find out is to build it and stress-test it yourself.

Python PyTorch vLLM TensorRT Triton FastAPI Docker LangGraph LangChain

Voice AI Agent

• AI Call Center Voice Agent — Outbound AI call bot + inbound complaint intake. 5-stage hybrid routing reduced LLM calls by 85%, GPU utilization 2.3% per call
• Multilingual Simultaneous Interpretation & Voice Analysis — 3 models (ASR+Translation+TTS) co-served on single GPU, 13 languages. Correctional facility risk utterance detection (7-stage NLP pipeline, sub-ms)

STT/TTS Model Evaluation & Optimization

• STT 3-Model Benchmark — Zipformer2 vs Qwen3-ASR under 200 concurrent / 30min sustained load (RPS 1,036, p95 250ms)
• ASR TensorRT Optimization — Qwen3-ASR encoder TRT FP16: inference 11.2ms → 5.6ms (2x), model size -49%
• TTS 4-Engine Comparison — CosyVoice2/MeloTTS/Qwen3-TTS evaluated on RTF/TTFA/concurrency/VRAM under same GPU. Built 3-stage defense for GPU Stuck recovery
• Finetuning Effect Quantification — CosyVoice2 Base vs ft_ko: RTF +32%, TTFA -61%, degenerate outputs eliminated

Production Systems

• English Education ASR System — Triton + Faster Whisper, 3-Worker architecture. Success rate 82% → 95%+, 300 req/min stable on 2x RTX 3090
• Enterprise Document RAG QA — LangGraph dual-graph design, Upstage Document Parse, recursive Map-Reduce summarization
• Custom Voice Keyword Spotting — Dual-threshold detection: 96.81% recall, 0.0% false alarm (430K+ window test). TFLite INT8 edge deployment
• VoiceNote Meeting Analyzer — Whisper STT + pyannote diarization + LLM summarization, 5 microservices
• Senior Care Chatbot — LLM-based elderly health check, 2-level state machine (16 sub-states, 40+ transitions)

AI R&D

• WIGVO — Real-time bidirectional speech translation over PSTN. Dual-session + 3-stage echo filter, 0 echo loops in 147 live calls, 557ms latency, $0.27/min (ACL 2026 Demo Accepted, 2nd Author, Rating 7.50) Repo
• WigtnOCR — VLM-based Korean public document parser, 2B model matching 30B Teacher quality Repo

Open Source

• WIGTN-Coding — Claude Code AI-Native development workflow plugins Repo
• TimeLens — Multimodal AI museum curator with voice + camera (Gemini Live Agent) Repo

Products

• WIGVU — YouTube subtitle extraction, translation & AI analysis service Repo
• WIGENT — AI Agent real-time debate platform 🏆 Build with TRAE Hackathon Grand Prize

• 🏆 Build with TRAE Hackathon · Grand Prize · WIGENT: AI Agent Real-time Debate Platform · 2026
• 🏆 International oneM2M Hackathon · Encouragement Award · 2022
• 📝 Implementation of an IoT Cocktail Machine Using ChatGPT API and ConvAnalyser in the Metaverse (IEEE Metacom 2024)
• 📝 A Metaverse Avatar Teleport System Using an AIoT Pose Estimation Device (IEEE Metacom 2023)
• 📄 IoT-based Metaverse Management Platform · Patent 10-2023-0189803