Yunbeom Choe

MSc Data Science & AI @ University of Liverpool. Interested in Generative AI, MLOps, and Reinforcement Learning. How I got here →

Hi, I’m Yunbeom Choe

MSc Data Science & AI student at the University of Liverpool, with a background in naval architecture and ocean engineering (Inha University). The shift from modelling physical systems to building ML pipelines was less of a leap than it sounds — both come down to systematic experimentation and honest failure analysis. I wrote more about the transition here.

Projects

DefectVision - Manufacturing Defect Detector

Real-time manufacturing defect detection using unsupervised anomaly detection — trained on normal images only, no labeled defects required.

  • Stack: Anomalib, PatchCore, PyTorch, OpenVINO, FastAPI, Streamlit, OpenCV, Docker
  • 100% Image AUROC on MVTec AD bottle; 8–20pp drop on the harder MVTec AD 2 benchmark
  • Lighting augmentation experiment: discovered augmentation hurts memory-bank methods by widening the normal distribution in feature space
  • FastAPI inference API with /calibrate endpoint for on-site threshold tuning
  • Real-time webcam streaming with queue-based non-blocking inference pipeline
  • 14 tests, CI/CD with GitHub Actions, Docker deployment

Read the post

FinScope - Multi-Agent Financial Report Analyst

Multi-agent RAG system that analyses SEC EDGAR and Companies House filings through a 3-agent pipeline.

  • Stack: LangGraph, Groq (llama-3.3-70b), ChromaDB, FastAPI, Streamlit, Docker
  • Retriever → Analyzer → Critic pipeline with conditional retry loop
  • Parallel Risk / Growth / Competitor analysis via asyncio.gather
  • Hybrid retrieval (dense + BM25 + RRF + cross-encoder rerank)
  • Critic agent: LLM-as-judge hallucination check with fail-open design and retry guard
  • Extended from arXiv RAG - same retrieval core, new multi-agent orchestration layer

Part 1: Building the System · Part 2: Critic Eval

arXiv RAG System

End-to-end Retrieval-Augmented Generation system for querying academic papers from arXiv.

  • Stack: FastAPI, ChromaDB, Qwen3 4B (via Ollama), Streamlit, Docker
  • 7-day build: broken embedding pipeline on Day 1, systematic retrieval optimisation that hit 100% hit rate by Day 5
  • Hybrid retrieval (dense + sparse) + cross-encoder reranking over 153 arXiv papers
  • LoRA fine-tuning experiment — documented a 28pp regression caused by training data contamination
  • Async refactoring of the entire I/O pipeline (FastAPI + httpx), fixing 7 bugs in the process
  • Fully local inference on Apple M4 Pro via Ollama — no external API calls

Part 1: Building the system · Part 2: Async refactoring · Part 3: LoRA Fine-Tuning

TORCS Corkscrew RL Racing Agent

Autonomous racing agent trained on the TORCS Corkscrew track using deep reinforcement learning.

  • Stack: Python, Stable-Baselines3, SAC, PPO
  • 9.7M training steps across 4,349 episodes — 37 track completions (0.85%)
  • Systematic failure mode analysis: 52.59% early crashes, 32.33% S-curve failures
  • Reward shaping, hyperparameter sensitivity analysis, and catastrophic forgetting investigation

Read the post

Skills

Languages: Python, SQL
ML/AI: PyTorch, Hugging Face, LangChain, LangGraph, RAG, Multi-Agent Systems, RL (SAC, PPO), LoRA fine-tuning, Anomaly Detection (PatchCore), OpenVINO MLOps: FastAPI, Docker, ChromaDB, pytest, GitHub Actions Tools: Git, Ollama, Streamlit

Contact