Tech Stack: Python, TensorFlow

Expert’s cases:

1. Developed a system to predict SMPL parameters from 3D point cloud data of the human body. Used the SMPL parameters for skeleton and individual bone predictions. Additionally, utilizing Graph Neural Networks to predict human body metrics such as height and weight

2. Working towards applying advanced Machine Learning techniques to analyze and predict injury risks based on comprehensive data sets, including point cloud data and other relevant information collected over a span of three years from a cohort of 430 athletes/runners at the university

3. Played a role in the information extraction from Safety Data Sheets (SDS), including training the BERT model for text representation learning, utilizing CatBoost for the classification of 16 sections, and showcasing adept problem-solving skills. This initiative resulted in a significant accuracy improvement from 0.81 to 0.89

4. Led a team of 4 in automating information extraction from real estate documents, resulting in substantial time and workforce savings for 7 clients. This included fine-tuning LayoutLM and leveraging large language models to minimize the need for manual intervention

5. Led a project focused on exact phrase matching for SDS. Implemented Faiss indexing for accelerated data retrieval, resulting in a 10-fold reduction in query time. Engineered a robust TF-IDF representation for precise word-to-word similarity, further enhancing matching accuracy. Additionally, designed and implemented an efficient API collaborating with frontend and data teams

6. Recognized the need for a tailored approach to different table types, shifting from a one-size-fits-all model. Employed classical image processing techniques for bordered table detection and extraction, achieving a TED score of 0.97, indicating near-perfect extraction. Additionally, utilized a Deep Learning-based image segmentation model for borderless tables, surpassing extraction performance compared to industry-standard Adobe and Microsoft tools with a TED score of 0.78

7. Developed a clothes size estimation system, resulting in improved models and a 50% reduction in pipeline latency