Harim Choi blog

WELCOME TO HARIM'S PAGE.

I am a machine learning engineer interested in vectorizing reality, making predictions, and estimating uncertainty.

Over seven years, I have worked with public-procurement, construction, and retail data, connecting tabular ML, NLP, computer vision, and LLM agents from data collection through evaluation and deployment. I quickly learn unfamiliar domains and inherited systems, then turn data, models, and operating workflows into working services. I build systems that balance predictive performance with reliability, cost, and operational constraints.

WarpQuant INT3 LLM quantization overview WarpQuant Technical report / low-bit LLM inference LLM Quantization · Efficient Inference A dual-domain post-training quantization method that compresses transformer weights after Hadamard rotation and spends a small recovery budget on columns ranked by Output-Fisher sensitivity. The same report evaluates a TurboQuant-style KV cache and per-token INT8 activations. Qwen3.8-27B · 3.6165 text bpw Read Aug 16, 2026 · 1 min Procurement NLP flow from a notice title through weak labels to an INT8 CPU API Procurement NLP Production NLP / weak supervision / CPU serving Korean NLP · Weak Supervision · CPU Serving A CPU NLP API that classifies bid eligibility and detailed work categories from Korean procurement-notice titles. Macro-F1 0.9639 · category F1 0.90 · about 50 ms CPU Read Aug 16, 2026 · 2 min Synthetic bimodal distribution comparing a collapsed cumulative interval with per-bin conformal regions that preserve the low-density gap between modes R2CCP Bid Prediction Probabilistic ML / conformal prediction Conformal Prediction · Multimodal Forecasting A multimodal bid-rate forecasting system that preserves disjoint conformal regions and turns eight context distributions into candidate decisions through 500K Monte Carlo simulation. 90.73% coverage · PQ bid KPI +35% Read Aug 16, 2026 · 2 min One MCP connecting search, web extraction, academic PDF parsing, and recovery google-surf-mcp Public OSS / MCP search infrastructure MCP · Search Infrastructure One MCP for general search, academic search, web extraction, and spatial PDF parsing, with parallel queries and recoverable CAPTCHA handoff. MCP TOPLIST · Top 1% of 95K servers Read Aug 16, 2026 · 2 min Automated PKM flow from sharing through organization to search Monogram Public OSS / personal knowledge automation PKM · Knowledge Automation A personal knowledge system that turns one share action into organized, searchable knowledge from news, social media, papers, videos, and documents. One share action · capture, organize, search Read Aug 16, 2026 · 2 min Agentic Browser flow from scoped proposal through approval to an auditable result Bau Browser Private local MVP / agent-native browser Agentic Browser · Agent Safety A local-first desktop browser where users control private data and agent actions through a scoped propose → approve → execute flow. New-domain tasks · exact action binding 36/36 · out-of-scope actions 0 Read Aug 16, 2026 · 2 min Bar chart showing that task-side features, 2026 routing methods, deployment bandits, and activation probes fail to identify the per-task winning agent configuration before execution Vargo Private research / reliable agent evaluation Agent Evaluation · Decision Science Research asking whether task text, embeddings, or hidden states can identify the best LLM-agent configuration before execution. Oracle-gap recovery 54% → 85% · TMLR review Read Aug 16, 2026 · 2 min Cumulative wealth and drawdown curves comparing a three-seed TQC ensemble with IVV across 756 held-out trading sessions from 2019 to 2021 EMH Agent Public OSS / point-in-time RL portfolio agent Reinforcement Learning · Portfolio Decision A reinforcement-learning investment agent that builds and evaluates portfolios using only the information available at each point in time. Cost-aware 3-year validation · TQC +183.9% vs. IVV +99.6% Read Aug 16, 2026 · 2 min MyShot tracking a golfer, club, ball, pose skeleton, joint angles, and the ball path in one swing MyShot Private research / 3D golf vision and shot prediction Object Detection · Human Pose · 3D Vision · Ball-flight Prediction Connects golfer, club, and ball tracking with 3D swing mechanics and personalized shot-outcome prediction from one phone video. Clubhead-speed median error 6.1% · MPJPE 35.6 mm · X-Factor MAE 3.1° Read Aug 16, 2026 · 3 min E-AT result summary on balanced four-class ISIC: macro-F1 0.8647 and minimum expected calibration error 1.03 percent at separate checkpoints EAT Private research / medical-image calibration Calibration · Medical Vision E-AT combines focal loss, bidirectional R-Drop, and adaptive FGM so an ISIC skin-lesion classifier learns both class quality and better-calibrated confidence. ECE 1.03% vs. CR-SAM (AAAI 2024) 1.7% Read Aug 16, 2026 · 2 min Bar chart comparing WSSS performance of WeCLIP+ at 51.8 percent mIoU and the refined method at 53.31 percent mIoU on all 40,137 COCO-Val 2014 images WSSS Commissioned state-of-the-art weakly supervised segmentation research Weak Supervision · Semantic Segmentation Improves segmentation learned from image-level labels by finding unreliable pixels in automatically generated masks and repairing only those regions before retraining. Commissioned WSSS SOTA research · 53.31% mIoU · +1.5 pp over WeCLIP+ Read Aug 16, 2026 · 2 min Travel Agent flow from a natural-language request to three packages and CRM handoff LangGraph Travel Agent Delivered to a U.S. travel agency / Public OSS Agent Workflow · API Orchestration A travel advisor delivered to a U.S. agency that turns a natural-language request into three bookable packages and connects review, CRM, and customer messaging. Agent delivered to a U.S. travel agency Read Aug 16, 2026 · 2 min