Machine Learning Deep

NIHAAL CHOWDARY SURPANI

Ó +1-302-***-**** R *************@*****.*** linkedin Github Website

EDUCATION

University of Delaware Newark,DE

Master of Sciences in Data Science, GPA: 3.8/4 February 2023 - December 2024 Courses: Imaging and Deep Learning, NLP, Machine Learning, Independent Research (worked on Transformers), Databases, Big Data Analytics, Multivariate Statistical Analysis. KL University Vijayawada, INDIA

Bachelor of Technology, Electronics and Communications, July 2018 - May 2022

WORK EXPERIENCE

University of Delaware Newark, Delaware

Machine Learning Researcher (CRP Lab) August 2023 - Present

• Optimized FLAN-T5 with precise fine-tuning using a hate speech reward model, applying PPO and RLHF to achieve a 15% performance boost in generating positive summaries from the text.

• Leveraged PEFT for efficient fine-tuning, reducing training time by 35%, while maintaining model performance.

• Designed a Machine learning model for flocculation size prediction in seawater with a 1- hour forecast horizon.

• Achieved 92% accuracy and 100x reduction in computational time compared to traditional numerical models.

Graduate Research Assistant (CIML Lab) December 2023 - June 2024

• Developed a novel Hierarchical transformer architecture and improved hyper-spectral image (HSI) reconstruction accuracy by 40%, outperforming traditional deep learning algorithms.

• Created a mask-guided mechanism to generate high-quality 19-channel image from a single snapshot, enabling more detailed spectral analysis within the 400 to 700 nm wavelength range.

• Reduced computation costs by 60% through the integration of a window attention mechanism.

Center of Atmospheric Sciences, KL University Vijayawada, India Research Assistant July 2021 - July 2022

• Pioneered a Deep Learning model to forecast wind power potential and speed. Implemented a cascading workflow that improved prediction accuracy by 10% (Publication) (Code).

• Enhanced wind turbine utilization by 50% through accurate wind speed predictions, deployed the model on AWS with EC2 and Flask for fast and reliable responses.

PROJECTS

QUESTION ANSWERING SYSTEM [TensorFlow, PEFT, LLM, Pipelines]

• Optimized DistilBERT by fine-tuning on the SQuAD2 dataset for Question Answering task, achieving 85% accuracy and enhancing performance (Code).

• Employed PEFT and LoRA techniques to enhance model fine-tuning processes, achieving a notable 40% reduction in training time while ensuring optimal model accuracy.

PINHOLE CAMERA [Pytorch, cv2, U-Net, NumPy]

• Designed Autoencoder model that enhances clarity of images captured through pinhole-sized aperture by 5 times.

• Developed a comprehensive dataset of images captured using two different aperture sizes, enhanced dataset clarity by applying advanced image processing techniques, and reduced image noise by 30% (Code). SKILLS

Programming Languages: Python, SQL, SAS, R, C, C++, Java, HTML. Technologies and Frameworks: TensorFlow, PyTorch, Keras, Scikit-learn, Seaborn, OpenCV, NLTK, NumPy, Pandas, Streamlit, FastAPI, Git, CI/CD, Docker, Flask, CUDA, HPC, AWS, Power BI, JMP, LangChain, Llamaindex MLflow. Core Skills: Machine Learning, Deep Learning, Natural Language Processing(NLP), Computer Vision, Data analytics .

LEADERSHIP AND ACHIEVEMENTS

• Led a team of 4 to optimize an open-source tool for Chemours by automating the chemical testing process, reduced overall research time by 20%, and secured first place in the University of Delaware DSI Hackathon (2023).

• resolved 60+ critical code issues within AMPL software, improving performance and robustness, leading to a consec- utive victory in the 2024 hackathon (Released Version).

• AWS certified cloud practitioner (EC2, S3, VPC, Sagemaker, lambda, RDS).

Contact this candidate