Seoul

Harim
Choi.

Production ML engineer, 6+ years end-to-end across CV, NLP, predictive. Self-taught, non-traditional path. Ship cycles: MCP server in 2 days, Monogram in 2 weeks, bidNLP and langgraph in 1 month each, R2CCP in 2 months, WSSS in 3 months (SOTA at release). All projects in 1+ year production unless research.

Production ML + OSS + research.

github → linkedin → email →

Python · TypeScript · PyTorch

MCP · LangGraph · Agent infra

Conformal · Calibration · Weak supervision

ONNX INT8 · FastAPI · GitHub Actions

wsss-refined-pseudolabels

56.2% mIoU · SOTA at release · 3 mo

Weakly-supervised semantic segmentation, refined pseudo-labels. Frozen CLIP (ViT-B/16) + DINOv2 backbone, Multi-Signal Reliability Estimation, Phase-Adaptive Refinement, Boundary-Aware loss. 56.2% mIoU on COCO-Val: +4.3pp over WeCLIP+ 80K baseline, +9.1pp over WeCLIP (CVPR 2024). SOTA at release. Delivered in 3 months. External contracted research, independently secured.

PyTorch · CLIP · DINOv2 · WSSS github →

nlp-analysis-agent

F1 96.4% · 50 ms CPU · 1+ yr prod

Korean public-procurement notice classification (bidNLP). RoBERTa-large + LoRA Teacher-Student, UltimateTrainer (FocalLoss + R-Drop + FGM). Hybrid weak labels: hard-rule routing + SBERT (0.9) / finetuned-RoBERTa (0.1) max-sim ensemble. Static INT8 ONNX (LoRA-merge → AVX512-VNNI per-tensor, 200-sample MinMax calib): 1.3 GB → 330 MB, 150 ms → 50 ms, <1% F1 loss. FastAPI service, 1+ year in production. Processing 70,000 notices/week: 40 hr manual → 2 min automated. F1 96.4% vs GPT-4o 35.1% (2.75×).

Python · RoBERTa · ONNX · FastAPI github →

monogram

PyPI · mono-gram

Drop into Telegram. Auto-save as wiki. Wake up to a project dashboard. 5-stage LLM pipeline, atomic Git Tree commits, MCP server (13 tools).

Python · LLM · MCP github →

google-surf-mcp

209 stars · 27 forks

Vendor-agnostic Google search MCP server. Drop-in for Claude Desktop, Cursor, or any MCP-compatible client. SSRF-hardened, 11 test cases, npm-published.

TypeScript · MCP · npm github →

ensemble-bid-prediction

+25-40% win rate

R2CCP for tender bid rate prediction. Identified interval collapse in the public implementation (cumulative-mass intervals merge bimodal peaks). Fixed via per-bin threshold + entropy regularization → bimodal preserved. +25 to 40% bid win rate, 1+ year deployed.

Python · R2CCP · Conformal github →

langgraph-travel-agent

7 modules · 4 APIs

Multi-agent travel booking with 4-API parallel orchestration and human-in-the-loop checkpointing. Refactored 1707-line monolith into 7 clean modules.

Python · LangGraph · HITL github →

claude-setup

dotfiles

Personal Claude Code configuration. Hooks, slash commands, statusline, MCP servers, project-level CLAUDE.md priors. Reproducible across machines.

Shell · Hooks · MCP github →

recently

Active across three tracks in parallel.

production: bidNLP: F1 96.4%, 50 ms CPU, 1+ year in production. R2CCP custom impl: +25-40% win rate, 1+ year deployed. NGBoost × XGBoost bid price: win rate 3× baseline.
research: DSSP: 12-branch decision-science taxonomy for LLM agents, 14 audits, arXiv preprint coming. E-AT: entropy-based adversarial calibration, LAPC loss family v1-v7.
agent infra: monogram: 5-stage LLM pipeline + 13-tool MCP server, atomic Git Tree commits. google-surf-mcp: vendor-agnostic Google search MCP, 209 stars (141 in first 5 days), npm-published.