I specialize in Large Language Models (LLMs) and Natural Language Processing (NLP), with a focus on advancing the state-of-the-art in downstream tasks such as semantic understanding, timeline summarization, and grammatical error correction. My work bridges cutting-edge research (4 ACL publications) and production-grade systems (ByteDance, Apple), making AI accessible and practical. Last but not least, my first name "Geonsik" is pronounced as "Gun-Shik" or /kʌn.ɕik/. You can also just call me "GS" 🙋🏻‍♂️

• Fine-tuning & Optimization: LoRA-based PEFT, 4-bit quantization, hyperparameter tuning with W&B
• Model Deployment: vLLM serving, LangChain pipelines, ChromaDB integration, AWS Bedrock inference
• Models: Mistral, Llama 2/3, FLAN-T5, GPT family, OpenAI GPT-OSS-20B
• Structured Outputs: Using instructor library with Pydantic models for reliable LLM responses
• Research: Incremental clustering algorithms using LLM-based pairwise classification

• Grammatical Error Correction (GEC): Sequence-to-sequence & sequence tagging approaches
• Timeline Summarization (TLS): Event detection, clustering, and narrative construction
• Semantic Understanding: Word Sense Disambiguation (WSD), Words-in-Context (WiC)
• Email Classification: Topic-based email classification with RAG-enriched semantic search
• Transfer Learning: Encoder-only vs. decoder-only architectures for semantic tasks

• Scalable Web Applications: Flask, Streamlit, Bootstrap, Docker containerization, LAMP stack
• Microservices Architecture: GEC system with separate API and web interface modules, email processing pipelines
• Model Serving: Production-grade deployment of transformer models and LLMs for real-time inference
• RAG Pipelines: Retrieval-Augmented Generation with vector embeddings for semantic search

• Novel approach leveraging LLMs for incremental event clustering and timeline construction from text streams. Outperformed SOTA on 4 TLS benchmarks.
• Tech Stack: PyTorch vLLM Llama-2-13B LangChain ChromaDB

• Comprehensive framework demonstrating encoder-only models outperform decoder-only LLMs on word meaning comprehension tasks.
• Tech Stack: PyTorch HuggingFace Transformers LoRA PEFT WandB

• End-to-end email processing pipeline with Streamlit web UI for Gmail integration, intelligent topic classification using AWS Bedrock LLMs, and AI-generated email thread summaries. Features incremental processing, LLM-powered topic attribute generation, RAG-enriched classification with semantic search, and complete lifecycle management (create, view, delete projects).
• Tech Stack: Python Streamlit AWS Bedrock OpenAI GPT-OSS-20B instructor Pydantic Gmail API FAISS Amazon Titan Embeddings ChromaDB LangChain

• Research toolkit for analyzing LLM agent trajectories on software engineering tasks.
• Tech Stack: Jupyter Python Agent Frameworks

• Self-contained archive of LeetCode solutions demonstrating strong algorithmic foundations.
• Tech Stack: Python Data Structures Algorithms

1. From Moments to Milestones: Incremental Timeline Summarization Leveraging Large Language Models
   Qisheng Hu, Geonsik Moon, Hwee Tou Ng
   ACL 2024 (Main Conference) | [Code | Paper]

2. Are Decoder-Only Language Models Better than Encoder-Only Language Models in Understanding Word Meaning?
   Muhammad Qorib, Geonsik Moon, Hwee Tou Ng
   ACL 2024 (Findings) | [Code | Paper]

3. ALLECS: A Lightweight Language Error Correction System
   Muhammad Reza Qorib, Geonsik Moon, Hwee Tou Ng
   EACL 2023 (System Demonstrations) | [Code | Paper]

4. WAMP: Writing, Annotation, and Marking Platform
   Geonsik Moon, Muhammad Reza Qorib, Daniel Dahlmeier, Hwee Tou Ng
   IJCNLP-AACL 2023 (System Demonstrations) | [Code | Paper]

• 🌐 Website: gsmoon97.github.io
• 💼 LinkedIn: linkedin.com/in/gsmoon97
• 🎓 Google Scholar: si3AXV8AAAA
• 🔬 ORCID: 0009-0001-5646-466X
• 📧 Based in: New York, NY