Boston University

Master of Science in Computer Science. Boston, MA.

National Yang Ming Chiao Tung University

Master of Engineering in Electrical and Computer Engineering.

University of Illinois Urbana-Champaign

Exchange Program.

National Chiao Tung University

Bachelor of Engineering in Electrical and Computer Engineering.

Engineer, Realtek Semiconductor Corp.

Owned development of a Python-based capacitor-coupling verification tool to automate design checks.
Coordinated API exposure and extension requirements with the internal design-tools team, documented tool usage, and brought relevant teams together for progress updates and feature discussions.
Designed and implemented a dual-frequency oscillator behavioral model for circuit analysis and validation.

StoryFlow: AI-Powered Visual Novel Editor

Built a Node.js visual novel editor with graph-based branching narratives, interactive scene previews, character staging, and AI-generated voiceovers.

Organized story graphs and asset workflows into modular components for rapid iteration.

ScriptToVideo: Script-to-Narrated-Video Application

Built the frontend, backend, and AI API integration from zero to one, delivering an internet-accessible workflow from script parsing to narrated video.

Persisted user assets and generation stages in Firebase across browser restarts.

Implemented OAuth2 authentication and atomic credit reservations, charging on successful generation and refunding failures.

Identified provider API throughput as a constraint on concurrent usage.

Sticker Gacha: Digital Stamp Collection System

A standalone collection app for Chibi stickers featuring a Gemini AI-powered Gacha system and infinite expression generation.

Generates images from a curated pool based on given character, topic, and style references, all within a premium shimmering UI.

CMV (Change My View) Essay Tone Categorization

Built a staged LangGraph workflow that answers one question at a time, grounds answers in sentence-level evidence, and records supporting rationale.

Evaluated 100 generated essays using human review of a subset and weaker-model comparisons against a stronger reference model, reaching 90% agreement with the reference.

DebiasPI: Inference-Time Debiasing by Prompt Iteration

AIEM Group, supervised by Professor Margrit Betke.

Developed a prompt-iteration framework to steer demographic distributions in DALL-E 3 outputs.

Designed an image-annotation codebook and automated bias-measurement pipeline.

Reached assigned race distributions while documenting persistent skin-tone imbalance and limits of prompt-based control.

Text Guided Image Integration

Proposed the project approach and aligned teammates on experiments and evaluation.

Prepared 7.6K training and 0.6K evaluation examples from video frames, character segmentation, background inpainting, and scene descriptions.

Fine-tuned ControlNet for text-guided character placement and evaluated similarity, text alignment, and realism; investigated limitations in dataset size and quality.

PromptsHub: AlphaMind Edition

Engineered a high-performance workspace for AI-human collaboration featuring real-time screen OCR, vision chat (Gemini Vision), and a multi-pane development environment.

Implemented a 3-column "Monitor Studio" for persistent region monitoring, system audio transcription with FFT visualizers, and a privacy mode to exclude workspace captures.

SlideTalk: AI-Powered Presentation Narration

SlideTalk is a powerful AI-assisted tool that utilizes Google's Gemini models to automatically generate professional narration, voiceovers, speaker scripts, and notes for your presentation slides.

Utilizing Range Deletes as a Query Filter in LSM-Based RocksDB

Targeted inefficiencies in RocksDB point queries by proposing and prototyping level-wise and SuRF-based range-delete filters, reducing unnecessary disk I/O and accelerating query speed.

Profiled bottlenecks with Valgrind and Intel VTune and implemented range-tombstone propagation during flush and compaction.

Achieved up to 40× higher throughput for queries on range-deleted keys while preserving baseline performance for unaffected workloads.

Built a verifier class and scripted regression checks to validate behavioral correctness and log errors after implementation changes.