A model that looks perfect is usually leaking somewhere. Finance taught me that suspicion, and it's how I work now.

I came up in crypto-finance — research and trading at a hedge fund and VC, where I analyzed projects, took positions, and built the Python bots my desk actually ran on: real-time alpha tracking, anomaly and gap detection, signal monitoring. I wrote 300+ research reports along the way. When being wrong costs you money, you learn to distrust data that looks too clean.

I bring that same habit to the factory floor now — working with real-time data, optimizing for cost, and pushing on a model until I'm sure it actually holds. Manufacturing ML, energy optimization, agentic pipelines, and systems that actually ship.

I don't trust a model until I've tried to break it.

• Industrial AI — defect detection, predictive quality, energy & cost optimization, risk scoring
• Data Engineering — ingestion-to-serving pipelines, on-prem & cloud, with audit-ready lineage
• Agentic & real-time AI — MCP agents, local / air-gapped LLMs, streaming anomaly detection

MCP + a local LLM, fully air-gapped and audit-ready (IATF 16949 / 21 CFR Part 11). The AI writes the plan, deterministic code executes it, and a human approves.

• Designed the project skeleton and the end-to-end agentic-flow backbone (Inspector → Planner → Executor → Validator)
• Built the React 6-stage pipeline UI; led troubleshooting and limited-spec (8 GB VRAM) LLM benchmarking

Vision + structured ML fused into a risk score, served through a deployed 4-Layer MES on AWS. I built the entire ML stack and the React dashboard.

• YOLOv11s defect detector — Recall 0.9975, mAP@0.5 0.9875 (recall-first for safety-critical inspection)
• Caught target leakage (a fake AUC 1.0) and reported the honest 0.556 — proving structured data alone can't predict failure, which justified the vision track
• 100-point risk score ranking 251 detections from CRITICAL → LOW for the factory floor

Cut electricity cost and CO2 by optimizing power factor and peak load over 12 months of steel-process time-series — a surrogate model + optimizer + ROI loop.

• My role: feature engineering, model optimization, backend / DB
• Results: −18.5% annual CO2e · ₩4.41M cost saving · peak demand 660 → 590 kW
• Team MaD Fairies (3-person)

A Transformer-based detection system — the evolution of the trading-desk bots I built (surge / gap / anomaly detection) into a general real-time pipeline. Phase 1 (Transformer modeling) complete.

ML & Data Science

Python pandas scikit-learn LightGBM XGBoost PyTorch SHAP

Computer Vision · LLM & Agents

Ultralytics YOLOv11 MCP Ollama Gemma FastAPI

Data · Infra

PostgreSQL DuckDB Parquet Docker AWS GitHub Actions

Frontend

React TypeScript Vite

Crypto hedge fund & VC (AUM 10ドルM+).

• Researched token valuations (FDV/MC, tokenomics, investor structure, vesting) → strategy planning and position execution; published 30+ VC-grade research reports
• Interpreted on-chain & derivatives signals (OI, funding rate, liquidation maps, whale flow) and applied them to trading strategy
• Designed and built Python automation bots — Long/Short Signal Bot, Alpha Tracker, CME Gap Detector — for real-time alpha monitoring
• Contributed directly to hedge-fund long/short position adjustments

• Blockchain & capital-market research; statistical analysis of disclosures and market data to surface opportunities and support decisions
• Authored 300+ daily / weekly / monthly market reports
• Ran investment content channels (YouTube + Telegram, 27K / 5K subscribers), building brand and credibility
• Financial data processing & visualization (Excel, Google tools); OKR-based goal setting

• 🎓 B.A. in Business Administration, Dongguk University
• 🧪 Samsung AXI K-Digital Training (AI & Data) · Schneider Electric Academy (semiconductor / battery industrial solutions) · Build with AI Seoul 2026 (with Google DeepMind)
• 📜 ADsP · Engineer Big Data Analysis (written) · Certified Investment Manager (투자자산운용사)