High-Throughput Diffusion Inference System
Production inference system for a 12B-parameter Diffusion Transformer on H100 clusters — 20–30 images/min/GPU at 1–3s latency with multi-GPU NCCL coordination
Portfolio
Video Super-Resolution Pipeline
ESRGAN-based 360p→720p video upscaling system — avg PSNR 31.5 dB, SSIM 0.919, processing at 7.8 fps on GPU
AI Virtual Try-On System
Automated diffusion-based garment-swap pipeline — async task orchestration, smart polling, and batch processing across model×garment combinations
Multi-LLM Evaluation Platform
Cross-platform Electron desktop app with custom browser automation for side-by-side LLM evaluation — supporting a first-author ICML 2026 submission (31K+ trials, 5 LLMs, 1,989-task benchmark)
LLM-Powered Recommendation Platform
Full-stack RAG chatbot on GCP — BigQuery vector search (ScaNN), Gemini 2.0 Flash, LangChain — MRR 0.88, nDCG@3 0.85 across 500 queries on 571M Amazon reviews
Privacy-Preserving Ed Discussion RAG System
End-to-end Python pipeline transforming historical course forum data into RAG corpora, powering a grounded TA chatbot via hybrid retrieval (Dense + Lexical + Metadata) and Gemma-3-4B-IT.
Urban Microclimate Forecasting System
Multi-scale LSTM with attention and gated feature selection for campus-scale microclimate prediction — Temperature RMSE 0.43°C / R² 0.993, Relative Humidity RMSE 1.28% / R² 0.995, mapped at 100,283 grid points via Regression Kriging