Research in progress
Research in progress
Oct 2026–presentEvidence-Linked Memory and Document Retrieval
Built a Python retrieval benchmark across 1,500 public-data questions, comparing keyword, dense, hybrid and graph-based methods with source attribution, latency profiling and failure analysis; implemented versioned retrieval contracts and 20 functional checks.
1,500 questions · reproducible evaluation
Project notes+
Python, BM25, E5, RRF, PPR, PyTorch
Evaluated held-out HotpotQA and LoCoMo questions with source-level error analysis. Retained the stronger baselines after graph candidates underperformed, and documented the results for subsequent experiments.
Research in progress
Oct 2026–presentClinical Dialogue Context Evaluation
Reproduced a 15-stage lab LLM classification pipeline and built a six-condition context study over 241 unique utterances. Added paired error analysis and a Python/HTML review interface with sentence-level audio and HTTP Range seeking.
6 context strategies · 241 unique utterances
Project notes+
Python, Qwen3 14B, scikit-learn, HTML, HTTP Range
The context study and audio-review interface extend an existing lab platform. Evaluation uses development sessions; external validation is future work.
Systems and hardware
BRISK — Llama 3 Attention on FPGA
Co-authored a single-owner K/V residency design for exact Llama 3 attention on Alveo U50; member of the team awarded 1st Prize in the FPT 2026 Design Competition, Track B.
FPT 2026 Track B · 1st Prize
Project notes+
C++ HLS, Alveo U50, Vitis, XRT, Python
Team project focused on exact attention computation, with single-owner K/V residency designed to reduce data movement. Contributed as a co-author of the accompanying paper.
Academic project
Sep–Oct 2026RV32IMA Core and Configurable Data Cache
Built a five-stage RV32IMA core and configurable write-back cache in Chisel; verified 40 GEMM/cache configurations, 810 cache checks and 58 upstream ISA binaries using an independent ISA simulator and retirement lockstep.
40 configurations · 58 ISA tests
Project notes+
Chisel, Scala, Verilator, C++, Python, Docker
Implemented forwarding, hazard handling, write-back/write-allocate caching and tree-PLRU replacement. Verification is simulation-based; further cache performance analysis remains in progress.
Virtualization and Live Migration Lab
Automated a nested-KVM live-migration lab with QEMU, NFS and QMP; migrated a running Linux workload with 14 ms QEMU-reported VM downtime and built reproducible CPU, network and VM/microVM benchmarks.
Running-workload migration · reproducible lab
Project notes+
QEMU, KVM, NFS, QMP, Linux, Python, Bash
Measured 14 ms of QEMU-reported VM downtime in a controlled nested-KVM run. Benchmarks used perf, sysbench, iperf3 and a matched CNN workload.
Systolic Array Deep Learning Accelerator
Developed and verified hardware modules for a systolic-array deep learning accelerator on FPGA using Verilog, FPGA development tools and GTKWave.
Hardware design and verification
Project notes+
Verilog, FPGA, GTKWave
Implemented hardware modules in Verilog and inspected simulation waveforms in GTKWave to verify the accelerator design.
Neural Network Graph Analysis
Created reproducible PyTorch-to-ONNX export and analysis tools for ResNet-50, AlexNet and ViT-B/16; validated all three models and analyzed operators and parameter costs for hardware mapping.
Three models validated with ONNX checker
Project notes+
PyTorch, torchvision, ONNX, Netron
Handled ViT export by decomposing the unsupported fused attention path into standard operators, then validated the exported graphs with the ONNX checker.
AI and data mining
Wattbot 2025 Sustainability RAG
Built a document-grounded sustainability question-answering system with source traceability and prompt engineering; ranked 16th among 66 teams in Wattbot 2025.
16th / 66 teams
Project notes+
RAG, LLMs, prompt engineering
Focused on connecting generated answers to supporting source material.
AI CUP Bank Alert Account Prediction
Designed an XGBoost machine learning pipeline for bank alert account prediction; finished in the top 20% of 790 teams in AI CUP.
Top 20% / 790 teams
Project notes+
Python, XGBoost, machine learning
An AI CUP competition implementation using boosted decision trees.
Document Question Answering with RAG
Implemented automated document question answering with LangChain, Llama 3.2 and ChromaDB, combining semantic retrieval with language-model generation.
Document retrieval and generation
Project notes+
LangChain, Llama 3.2, ChromaDB
A vector-store retrieval pipeline connects source documents to generated responses.
Social Emotion Classification
Fine-tuned RoBERTa for six-class emotion classification with stratified validation and macro-F1 checkpoint selection; compared models, preprocessing, class weights and learning rates.
Six-class NLP experiment
Project notes+
PyTorch, Hugging Face, RoBERTa
The public notebook contains individual model-training code and an experiment report.
Reddit Text Mining and Visual Analytics
Built a text-mining workflow with TF-IDF, frequent-pattern mining and classifier comparisons; explored PCA, t-SNE and UMAP representations with interactive Plotly visualizations.
Text features, patterns and visualization
Project notes+
NLTK, scikit-learn, FP-Growth, Plotly
Course project covering Naive Bayes, decision trees and augmented pattern features, with individual implementation in the public notebook.
Multimodal Lecture Analysis
Extended a course scaffold with Pydantic transcript/video schemas and Gemini-based segment analysis, combining lecture transcription with visual insights.
Structured video and transcript analysis
Project notes+
Gemini API, Pydantic, Python
Extended supplied course code with structured transcript and video-analysis outputs.
Software and data systems
Smart Medical Patrol Vehicle System
Designed and implemented a hybrid spatial/graph database architecture using PostGIS and Neo4j, containerized with Docker. The project was selected by a Taichung clinic for a paid collaboration.
Paid clinic collaboration
Project notes+
PostGIS, Neo4j, Docker
The database architecture combines spatial information in PostGIS with relationships in Neo4j, packaged in a Docker environment.