Sandhya S. Email and Phone Number

• Contributed to research on Generative AI (Gen AI), specializing in Single Shot Few-Shot Prompting techniques.• Conducted comprehensive evaluations on open-source models such as Codellama 70B, Codellama 34B, and Mistral-70B Mixtral-70, Llama-70B, Claude v2, Sqlcoder-70B (2 and alpha), WizardCoder-33B and gpt (3.5 and 4). Assessed their accuracy in NL-to-SQL tasks using the Spider dataset and its evaluation script.• Investigated the impact of prompt length and dataset size on model performance (zero-shot prompting) across varying complexities of SQL queries (hardness).• Analyzed model performance under environments including Anyscale, Bedrock, self-hosted, and GPT platforms. Explored hosting options on cloud platforms like AWS.• Developed benchmarks incorporating cost, throughput, and tokens/inference to ascertain the optimal open-source model for NL-to-SQL tasks, aiming to substitute GPT-based solutions.

• Explored multiple regression and classification models, built a robust 2-stage model that predicts if a flight will be delayed on arrival and, if so, the delay time in minutes based on the weather data at departure.• Developed a system using US flight data to categorize delays (<30 min, <1 hour, etc.). Integrated Flask for Python ML scripts and HTML/CSS frontend. Incorporated weather data from APIs to enhance prediction accuracy.