Welcome to my portfolio
I research, build, and ship AI.
From model design to production pipelines.
After studying Architectural Engineering, I spent ten years on construction sites coordinating multi-stakeholder teams and managing large-scale projects. That experience translates directly into my work as an AI Research Engineer — drawing service boundaries in MSA architectures and making the architectural calls that complex pipelines demand.
I currently lead the Engineering Part within the BrainCrew Research Engineering team. At my previous company, I designed and built five AI systems solo over nine months, including an Advanced RAG Pipeline, an Admin Console, and a B2B SaaS Platform.
I also lead a developer crew called WIGTN, where we have open-sourced an AI-native development toolkit (WIGTN-Coding) and a VLM-based Korean public-document parser (WigtnOCR). Our real-time bidirectional speech-translation system WIGVO, which runs over the legacy PSTN audio band (G.711 μ-law, 8 kHz) and removes the echo loop in software alone via a novel Dual-Session Echo Gating architecture, was accepted to ACL 2026 System Demonstrations.
Professional history
Research Engineering Team / Engineering Part Lead
Founder & Lead
AX Business Division / AI Research Engineer
Housing Business Division / Architectural Site Engineer
Head Office / Construction Project Manager
Building Construction Division / Architectural Site Engineer
Research and community activities
GPU-Accelerated Laboratory / Research Member
Academic background
Bachelor's Degree, Architectural Engineering
Associate's Degree, Architectural Design
Training programs and certificates
Fundamentals of Accelerated Data Science
Applying AI for Research (Modulabs)
Prompt Engineering Program (The Prompt Company; Sujin Kang, Ph.D.)
Data & AI Alignment Engineer Program
Honors and recognitions
Team WIGTN — WIGTN FLAKE: a goal-driven neighborhood-intelligence platform
The user picks "what they want to do" — open a café, market rental appliances, place an outdoor ad, invest in real estate — and an AI panel of domain experts cross-analyzes four datasets (real-estate prices, foot traffic, card sales, telecom contracts) to recommend the top three neighborhoods, automatically detect anomaly signals, and forecast a six-month trend. Built on a Snowflake Cortex × GPT-4o hybrid orchestration.
Led the on-site validation of the "H-Silent" inter-floor sound-insulation system; contributed to obtaining Korea's MOLIT Grade-1 certification.
"A Study on Improving Organizational Engagement on Construction Sites through the Introduction of Mentoring."
National qualifications
Korea Data Agency (K-DATA, 한국데이터산업진흥원)
Company & Side Projects
An AI Sales Enablement Platform that analyzes customer documents to auto-generate and ship a tailored RAG pipeline via Docker. Cut new-engine deployment lead time from 2 weeks to 5 minutes; ran 9 MSA projects · 161 APIs · up to 99 concurrent pipelines.
LoRA-fine-tuned Qwen3-VL-2B for Korean public-document parsing, surpassing Qwen3-VL-30B on OmniDocBench. Top-tier Hit@1 / MRR@10 in retrieval among open-source parsers. Released as an open-source HuggingFace model.
An architecture that breaks the echo loop in low-bandwidth PSTN telephony. Validated 0/148 failures in field testing; accepted as ACL 2026 System Demonstrations (first author).
An LLM-serving benchmarking tool usable by non-engineers. Visualizes TTFT / TPOT / Goodput and runs a dual cross-validation. Received cold outreach from an ex-Google ML/AI engineer.
An AI-native development workflow plugin built on Claude Code. 7 commands · 14 agents · 7 skills · 4 hooks. Quality-gate auto code review and a 3-tier memory system push pipeline throughput up by 3×.
When the user picks "what they want to do" (open a café, market rental appliances, place an ad board, invest in real estate, etc.), an AI panel of domain experts cross-analyzes four datasets (real-estate prices, foot traffic, card sales, telecom contracts) to recommend top-3 neighborhoods, surface anomaly signals automatically, and forecast a six-month trend. Built on a Snowflake Cortex × GPT-4o hybrid orchestration by a four-person Team WIGTN.
An entry for Google's global Gemini Live Agent Challenge. Uses a dual pipeline of Gemini Live API + ADK to process voice and camera in real time. Four function calls replace the routing logic, intent classifier, and state machine.
Learning & Side Projects
Three AI tools for FPS game development unified under the MCP standard. Generates weapon-recoil trajectories from natural language (98.7% time saved) and 3D map whiteboxing (75% time saved). Deployed on Smithery.ai.
A real-time exercise-form feedback service that combines YOLO11-Pose joint extraction with a LangChain RAG system. Achieved 100% Context Precision / Recall and 82% Answer Relevancy on RAGAS evaluation.
Full fine-tuning of KoGPT-2 on 7,439 song lyrics for an emotion-driven Korean lyric model. Analyzes the sentiment of input keywords across 43 KOTE emotion categories and generates tailored lyrics.
In a 24-hour hackathon, built an end-to-end pose-matching game combining real-time YOLO11-Pose recognition with Unreal Engine 5. Scores cosine similarity against reference poses; 30 fps real-time processing and sub-50 ms API response.
An end-to-end pipeline that auto-summarizes finance videos from a YouTube URL. Whisper-large-v3 for STT, then Gemma-2 with QLoRA 4-bit quantization to generate the summary on a single GPU.