Vedant Kene

Applied INT4/INT8 post-training quantization to vision transformer models using PyTorch and NVIDIA TensorRT, achieving ~4x memory reduction with less than 1% accuracy degradation. Built GPU-aware compression pipelines automating the full quantization workflow (model loading, calibration, engine export, and benchmarking), reducing turnaround time from hours to minutes. Designed benchmarking scripts to evaluate latency/throughput/accuracy tradeoffs and tracked experiments using structured Git branching for reproducible research.

Built Python-based RAG pipelines and REST API endpoints serving an AI-powered grading tool for a React frontend, reducing manual assessment effort by 60%. Automated data pipelines using AWS Lambda and REST APIs, cutting manual reporting effort by 45%, and delivered Power BI dashboards for management decision-making. Designed MongoDB collections and schemas for student data, model outputs, and session history. Collaborated in an Agile sprint workflow using Git feature-branch strategy, with structured testing and versioned iterations.

Built Python-based RAG pipelines and REST API endpoints powering an AI grading assistant used with a React frontend, reducing manual assessment effort by ~60%. Developed automated data pipelines using AWS Lambda and REST APIs, cutting manual reporting effort by ~45% and delivering Power BI dashboards for management decision-making. Designed MongoDB collections and SQL schemas to store student data, model outputs, and session history, collaborating in an Agile sprint workflow using Git feature-branching strategy. Also contributed to deep learning inference support and engineering practices for reliable experimentation.