LLM/VLM Prompt Optimization and Evaluation Framework
Built a reusable framework for versioned prompts, datasets, automated evaluation, experiment comparison, and error analysis. Optimized image-relevance classification with GEPA and validated gains through online A/B testing.
Large-Scale Urban Gaussian Splatting Pipeline
Built an end-to-end 3D reconstruction pipeline for scenes over 1 TB and tens of thousands of images, including COLMAP, training, fine-tuning, block merging, monitoring, and GPU orchestration.
Retail Demand Forecasting System
Developed a hybrid ARIMA and optimization-based forecasting system for about 15 retail chains, reducing modified MAPE by 15 percentage points while supporting manual delivery-calendar constraints.