AI Research Engineer @ WIGTN

LLM agents that hold up in production, including tool-calling control and guardrails, multi-agent orchestration, retrieval-grounded evaluation, and agents that act over real-time voice.

• Retrieval-Conditional Parsing Score (RCPS): Choosing Document Parsers by Retrieval, Not by Appearance EMNLP 2026 Industry Track · First author · GitHub
• WIGVO: Real-Time Bidirectional Speech Translation over Legacy PSTN Calls via Dual-Session Echo Gating ACL 2026 System Demonstrations · Second author · Paper · Video · GitHub
• Implementation of an IoT Cocktail Machine Using ChatGPT API and ConvAnalyser in the Metaverse IEEE MetaCom 2024 · First author · Paper
• A Metaverse Avatar Teleport System Using an AIoT Pose Estimation Device IEEE MetaCom 2023 · Paper

WIGTN · AI Research Engineer · 2026.01 – Present

• Develop WIGVO, a production real-time phone translation service over PSTN, improving translation quality through the echo gate and VAD pipeline, operator speaker identification, and a noise-robust audio front end.
• Built call observability and evaluation: per-turn Langfuse tracing, a live pipeline monitor, and an LLM-judge translation-quality evaluation with self-validation.

Soundmind · Team Lead, AX Division · 2026.03 – 2026.06

• Led a real-time voice AI agent engine for insurance sales-compliance monitoring: rule-based guardrails for simple intents, LLM tool calling only for risky utterances, complaints, and answer reversals.
• 5 concurrent calls on a single RTX 3090 with a 9B model at 154 ms P50 / under 300 ms P99, 47% fewer input tokens, 55/55 edge-case QA scenarios passed.

Soundmind · Manager, AX Division · 2025.01 – 2026.02

• Batch STT pipeline (Temporal, Triton, Faster Whisper): decoupled STT from result callbacks, success rate 82% → 95%+, 200 req/min on 2× RTX 3090.
• Custom Korean wake-word model (KWT-3): dual-threshold detection, 0 false alarms over 430K+ windows, 96.81% recognition, 75% smaller with TFLite INT8.
• Document RAG with a LangGraph dual graph (up to 50 MB / 100 pages) and a meeting-analysis platform with Whisper, pyannote diarization, and speaker-transcript alignment.

AID Lab, Hanshin University · 2022.09 – 2024.12

• Government-funded R&D on multimodal (text + speech) depression classification, on-device behavior recognition, and edge AI for small businesses.

• 🏆 ByteDance Build with TRAE Hackathon 2026 · Grand Prize — WIGENT: AI agents debate a business idea and turn the outcome into a landing page.
• 🏅 OBA Weekendthon S1 2026 · Top 6 — MyunZy: spoken mock interviews with four interviewer personas that follow up on weak answers.
• Google Cloud Rapid Agent Hackathon — Custos: always-on AI code review bot for GitLab.
• H0: Hack the Zero Stack · Vercel × AWS — OpenSlot: zero-oversell ticketing infrastructure.
• Gemini Live Agent Challenge · Google — TimeLens: multimodal AI museum curator with voice and camera.
• 🏅 International oneM2M Hackathon 2022 (Nov) · Encouragement Award — travel logging with biometric data from IoT wearables.

• Convergence of IT Image Data Analytics, Hanshin University (2023.09 – 2025.02) · GPA 4.5 / 4.5
• IT Transmedia Contents, Hanshin University (2018.03 – 2023.08)

• IoT-Based Metaverse Management Platform · Patent application No. 10-2023-0189803 (2023.12)
• Metaverse Emotion Mapping System Using an AIoT Facial-Expression Device · Software registration No. C-2023-054487 (2023.11)
• LSTM-Based Walking Motion Recognition in Industrial Sites Using YOLOv7 Pose Estimation · Software registration No. C-2023-049743 (2023.11)