Skip to content
View HarimxChoi's full-sized avatar
8 followers · 4 following

Achievements

Achievement: Starstruckx2Achievement: Pull Sharkx2Achievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Starstruckx2Achievement: Pull Sharkx2Achievement: QuickdrawAchievement: YOLO

Block or report HarimxChoi

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
HarimxChoi/README.md

Harim Choi

Portfolio preview

Production ML engineer in Seoul. 6+ years end-to-end across CV, NLP, predictive. Self-taught, non-traditional path.

Portfolio: https://harimxchoi.github.io

Production ML + OSS + research.

Recent

  • bidNLP: Korean public-procurement notice classification. RoBERTa-large + LoRA Teacher-Student, hybrid weak labels (SBERT + finetuned-RoBERTa max-sim ensemble), static INT8 ONNX (AVX512-VNNI). FastAPI service, 1+ year in production. F1 96.4%, 50 ms CPU, weekly 70,000 notices: 40 hr manual to 2 min automated.
  • R2CCP custom: tender bid rate prediction. Identified interval collapse in the public implementation, fixed via per-bin threshold + entropy regularization. +25 to 40% win rate, 1+ year deployed.
  • wsss-refined-pseudolabels: weakly-supervised semantic segmentation. Frozen CLIP (ViT-B/16) + DINOv2, Multi-Signal Reliability Estimation. 56.2% mIoU on COCO-Val (+4.3pp over WeCLIP+ 80K baseline). SOTA at release. 3-month contracted research.
  • monogram: PKM agent system. 5-stage LLM pipeline + atomic Git Tree commits + 13-tool MCP server. PyPI as mono-gram.
  • google-surf-mcp: vendor-agnostic Google search MCP server. SSRF-hardened, 11 test cases, npm-published. 209 stars, 27 forks.

Working on

  • DSSP: 12-branch decision-science taxonomy for LLM agents. 14 agent audits. arXiv preprint coming.
  • E-AT: entropy-based adversarial calibration. LAPC loss family v1-v7. Active research.

Contact

Pinned Loading

  1. wsss-refined-pseudolabels wsss-refined-pseudolabels Public

    WSSS with refined pseudo-labels: WeCLIP+ baseline + RFM refinement + disagreement-aware self-training (SOTA 56.2% mIoU on COCO-Val)

    Python

  2. langgraph-travel-agent langgraph-travel-agent Public

    Production-ready LangGraph multi-agent system for travel planning. Async parallel orchestration across Amadeus, Hotelbeds, Twilio, HubSpot.

    Python 12 2

  3. google-surf-mcp google-surf-mcp Public

    Google search MCP. No API key. One MCP replaces search + fetch + academic-paper extractor.

    TypeScript 209 27

  4. ensemble-bid-prediction ensemble-bid-prediction Public

    R2CCP-based PQ bidding: 8 context models, Monte Carlo 500K sim, conformal prediction with per-bin threshold

    Python

  5. monogram monogram Public

    Personal knowledge pipeline. Telegram -> 5-stage LLM -> markdown vault -> encrypted dashboard. MCP-queryable.

    Python

  6. nlp-analysis-agent nlp-analysis-agent Public

    Korean public procurement NLP: RoBERTa+LoRA bidability + multiclass with weak supervision (SBERT+LoRA), ONNX INT8 deploy

    Python