At Amazon, I built reusable data preparation pipelines for time-series forecasting and evaluated models across datasets. I developed a residual correction module for zero-shot Chronos, improving WAPE by 6 percentage points and Bias by 12 percentage points.
At Samsung R&D Institute India, I designed and implemented a training, evaluation, and inference pipeline for a lightweight Vision-Language Model for visual question answering and captioning. The model achieved comparable performance with a roughly 76% reduction in size.
At Dreamalle Solution Private Limited, I developed LLM-powered chatbots and virtual assistants, and built, fine-tuned, and deployed BERT and XGBoost models. My projects include a Zero-DCE image enhancement model and a FastAPI and Streamlit assessment recommendation system.
