Giwon Shin
Machine Learning Engineer and Researcher
Education
University of Wisconsin-Madison
Expected May 2026 • Master of Science in Electrical & Computer Engineering - Research • Madison, WI- Thesis: MuSe: Packing-Aware Causal Sequence Modeling under Fully Homomorphic Encryption
- GPA: 3.86/4.0
Yonsei University
Aug 2023 • Bachelor of Science in Engineering • Seoul, South Korea- Major: Electrical & Electronics Engineering, Civil Engineering
- GPA: 3.67/4.3
- Honors Thesis: Human pose dataset enhancement and segmentation model improvement
- Honors Thesis: Mobile robot based PPE compliance monitoring
Experience
Graduate Researcher
Dec 2024 to May 2026 • Grigorios Chrysos Research Group - UW-Madison • Madison, WI- Proposed MuSe, a packing-aware causal sequence architecture for privacy-preserving LLM inference under bootstrapped CKKS Homomorphic Encryption, replacing self-attention with Polynomial Toeplitz Mixer and feed-forward layers with a Mu-layer to reduce CKKS-unfriendly nonlinearities.
- Built and evaluated end-to-end pure-FHE causal inference in the DESILO CKKS stack under outer- and inner-based ciphertext packing; achieved 12.05x faster cache-enabled generation from 5,639s to 468s/token while maintaining low hidden-state deviation.
Machine Learning Engineer
May 2024 to Aug 2024 • Hecto Co., Ltd • Seoul, South Korea- Implemented quantized fine-tuning pipelines for large language models with HuggingFace and PyTorch, reducing GPU memory usage by 50% and doubling inference throughput across production NLP workloads.
- Designed and trained a custom object detection model with 95% F1, deployed scalable inference APIs with FastAPI, and integrated vector search with Milvus and Qdrant for the CycleAI production service.
- Developed a multimodal LLM-based text-to-speech system for OneApp, enabling automated customer interaction workflows.
Undergraduate Researcher
Feb 2022 to Nov 2023 • Smart Infrastructure Lab - Yonsei University • Seoul, South Korea- Co-authored 5 conference papers and 1 journal publication on machine learning-driven infrastructure monitoring.
- Developed crack segmentation algorithms and a semi-supervised domain adaptation framework using label propagation to address data scarcity in participatory sensing environments.
Projects
MuSe: Packing-Aware Causal Sequence Modeling under Fully Homomorphic Encryption
Jan 2026 to May 2026 • Homomorphic Encryption & Efficient Sequence Models- Designed a packing-aware causal decoder for pure-FHE inference, replacing softmax attention with Polynomial Toeplitz Mixer and feed-forward layers with Mu-layers.
- Evaluated encrypted causal generation under outer- and inner-based CKKS packing in the DESILO stack, with inner cached decoding achieving a 12.05x speedup from 5,639s to 468s/token.
Inducing and Analyzing Hallucination in Diffusion LLMs
Sep 2025 to Dec 2025 • Hallucination & Diffusion Models- Investigated hallucination behavior in the LLaDA-8B diffusion language model using anchor-token perturbations.
- Built an automated evaluation pipeline across 5 datasets and 500 queries, and classified hallucination types.
- Found hallucination rates above 80% independent of anchor position, revealing diffusion-specific error propagation.
Hybrid Text-Vision Retrieval for Multimodal RAG
Sep 2025 to Dec 2025 • Retrieval Systems & Multimodal ML- Evaluated visual retrieval with VisRAG against text retrieval with BGE and E5 on document-image datasets.
- Demonstrated text retrieval outperforming visual retrieval by 4x MRR on text-dominant documents.
- Proposed hybrid modality fusion that improved retrieval to MRR 0.8379 and Recall@10 0.928.
Gaze-aware Proactive AI Assistance for Digital Article Readers
Jan 2025 to May 2025 • HCI, Eye Tracking & LLM Systems- Built a browser-based reading assistant combining webcam gaze tracking, PDF.js document localization, and LLM-generated paragraph explanations for ESL readers.
- Implemented gaze-to-document coordinate alignment, heatmap overlays, and paragraph-level assistance workflows to ground AI help in the user's reading context.
Extending Tensor Depth Efficiency to Tensor-Train and RNN-like Models
Jan 2025 to May 2025 • Theory, Tensor Decomposition & Sequence Models- Extended a tensor-analysis depth-efficiency result from CP/HT decompositions to tensor-train factorization and a sequential RNN-like architecture.
- Derived an exponential CP-rank lower bound almost everywhere and validated the trend empirically through experiments on randomly generated TT tensors.
PseudoPremium: Optimizing Visitor Experience at Universal's Islands of Adventure
Sep 2024 to Dec 2024 • Optimization & Operations Research- Formulated a mixed-integer linear program in Julia/JuMP to optimize theme-park itineraries under ride preferences, wait times, ride durations, and must-ride constraints.
- Analyzed tradeoffs among satisfaction, queue time, express-pass decisions, and seasonal crowding to generate personalized park-planning recommendations.
Juno Meets Jupiter: Recreating and Reimagining JunoCam Imagery
Nov 2024 to Dec 2024 • Scientific Visualization & Image Processing- Reconstructed raw JunoCam imagery in Mathematica through channel-wise parameter search and color remapping to match enhanced Jupiter renderings.
- Created stylized but structurally informative visualizations of Jupiter that combined scientific reconstruction with expressive atmospheric rendering.
Something Fishy This Way Comes: Fish Segmentation in Mathematica
Sep 2024 to Oct 2024 • Image Segmentation & Classical Vision- Built an interpretable Mathematica segmentation pipeline using HSB decomposition, threshold fusion, morphology, and connected-component filtering.
- Recovered stable fish area, centroid, and horizontal-length measurements despite glare, shadows, and background-connected artifacts.
Certifications
Skills
Programming Languages
Python, MATLAB, SQL
Machine Learning
Pytorch, Huggingface, Tensorflow, ONNX
ML Domains
Computer Vision, Natural Language Processing, Deep Learning, Neural Networks, RAG
Systems & Infrastructure
Docker, FastAPI, Git, llama.cpp
Databases
PostgreSQL, Chroma, Qdrant, Milvus