Hyeontae Kim
Data Science Co-op @ University of Waterloo
Knowledge atlas
A living map of the ML and deep learning concepts I study.
I read papers, write implementation notes, and rebuild important methods in PyTorch. This graph is the main interface to the knowledge base behind the portfolio.
No graph concepts match this view.
Experience
Projects
FastAPI
STT
vLLM
LoRA
Smart Air Purifier Voice Control
End-to-end voice-control system for an interactive air-purifier robot, converting speech into validated device actions with SenseVoice STT, a LoRA-fine-tuned 4-bit Llama-3.2-1B model, grammar-constrained decoding, and a stateful animated UI. Reached 91.6% audio-command accuracy within a 3 GB VRAM / 0.45 s on-device target.
Repository HuggingFace
vLLM
llama.cpp
FastAPI
Pydantic
VLM Serving Benchmark & Routing Platform
Benchmarking and routing platform for document extraction, running one-command sweeps across Transformers, vLLM, llama.cpp, and FP16-to-INT4 quantization to choose the cheapest backend that still meets accuracy and JSON-reliability SLAs.
Repository
PyTorch
CV
NLP
Transformers
Deep Learning from Scratch
Hands-on implementations across computer vision and NLP, covering CNNs, detection and segmentation models, BERT/GPT style transformers, translation, NER, and next-token prediction. The focus is on understanding architecture internals by rebuilding them directly in PyTorch.
RAG
LangChain
Chroma
VectorDB
TechCrunch Q&A with RAG-Powered LangChain
Question-answering system specialized for TechCrunch articles, using LangChain, OpenAI embeddings, and Chroma to persist document chunks and ground answers in retrieved source context.
Repository
Python
OpenAI API
Airflow
AI arXiv Daily Translator
Automated pipeline that fetches, filters, and translates AI-related arXiv papers into Korean on a daily schedule, then compiles the translated papers into accessible PDF documents.
Repository
Compiler
C++
CS 241
Building a C Language Compiler from Scratch
Built a compiler for a subset of C in C++, covering scanning, parsing, context-sensitive analysis, and code generation with algorithms such as simplified maximal munch and bottom-up parsing.