Hossein Zolfaghari

– Designed and developed a Video Analysis API that automatically extracts representative keyframes and generates comprehensive video descriptions via multimodal LLM, enabling automated quality assessment and content under-standing of videos at scale.. – Built a keyframe sampling pipeline using perceptual hashing and CLIP-based semantic deduplication to maximize visual coverage per token budget, cutting VLM inference costs by minimizing redundant frame submissions

– Architected a multimodal time-series forecasting model (96% accuracy) and a real-time sky imaging pipeline for cloud motion estimation, cutting prediction error by 15%. – Built and deployed two YOLO-based vision services for grape growth stage tracking in vineyard footage and raindrop artifact detection in fisheye cameras, directly feeding into precision agriculture and solar irradiance workflows. – Designed an AI-driven PV panel orientation system and an end-to-end CI/CD pipeline using GitHub Actions and AWS for reproducible edge deployment on Raspberry Pi, boosting energy capture efficiency across distributed sites.

– Developed a high-precision solar irradiance forecasting model by applying advanced time-series techniques and feature engineering. The approach improved forecast accuracy by 5% over the previous model. – Created a sky-state estimation model using multi-modal environmental and imaging data, achieving over 90% classification accuracy.

– Built AI recommendation system serving 150K+ users with personalized content delivery. – Designed NLP pipeline for document processing handling 10K+ documents daily.

– Applied stochastic modeling and uncertainty quantification to build predictive models, improving accuracy by 25%. – Conducted comprehensive sensitivity and sampling analyses (Sobol indices, LHS) to evaluate model robustness and parameter impact, contributing to two publications, [Paper one, Paper two].