Rutvik Deshpande

• Built predictive models using Random Forest, Linear Regression, and Artificial Neural Networks to predict dementia risk in older African Americans, employing imputation methods (MICE, MissForest, MICEForest) to handle 65% missing data. • Published research papers as part of a $7.4 million NIH-funded grant, advancing machine learning methods for neurodegenerative disease prediction; presented findings at the Alzheimer’s Association International Conference (AAIC). • Engineered a Python-based ETL pipeline to automate REDCap data ingestion into MySQL, including database design, access control, and structured loading for 400+ patient records.

• Developed a modular LoRA fine-tuning framework that reduced trainable parameters by 81%, enabling fine-tuning, deployment, and evaluation of 10+ seq2seq models across Hugging Face and Lamini for tasks such as summarization and classification. • Developed an automated data pipeline using Airbyte’s REST API and Python SDK to enable transfer of client data from any database to any destination (e.g., Google Cloud Storage, MongoDB), facilitating scalable data ingestion across the organization. • Built the core backend service that extracts defanged IOCs from unstructured sources (e.g., PDFs, blogs), maps them to MITRE ATT&CK tactics and techniques, and exposes this intelligence via REST APIs used by SOC analysts across client organizations.

• Created a high level literature review on Open Data used in ML/AI by conducting research review on 20+ research papers • Improved resource and time management by predicting crime in Los Angeles by performing Statistical Analysis and fitting a Linear Regression Model on Crime Data • Fashioned Box Cox transformation on the non normal data • Devised hypothesis testing on various parameters to find the parameters for best fit

• Reduced query resolution time by designing a Slack RAG chatbot using LangChain and prompt engineering for query automation. • Optimized Slack chatbot performance by utilizing Arize AI for real-time model monitoring, performing anomaly detection and using CallbackManager for clusters with a drift score close to -1. • Fine-tuned T5 small model using LoRA (PEFT method) for text summarization, achieving a 69% increase in ROUGE Score.

• Reduced customer data retrieval time by 30 minutes per cycle by optimizing ETL processes in Ab Initio using parallel ingestion component. • Saved 6 hours of task processing per week by reorganizing and automating tasks in Ab Initio through Unix shell scripting. • Collaborated in developing an ETL database model on Ab Initio GDE to handle new procedures for customer complaints, processing over 3 million rows improving data processing efficiency.

• Accomplished a 96.4% accuracy in predicting future clients by building and deploying a PyTorch-based deep learning model on AWS Beanstalk, resulting in a 40% reduction in marketing costs through targeted outreach. • Enhanced service recommendation system by utilizing K-Means and Hierarchical Clustering with 6000 attributes for marketing campaigns, resulting in cumulative marketing cost cuts by 40%.

• Designed a real-time prediction model by integrating data from Telematics, surveys, web scraping and APIs into an Oracle database. • Improved journey planning by performing EDA on unstructured data and applying ARIMA for time-series forecasting.

Working and testing of Brushless DC Motors to generate power using Solar Energy.