I am a graduate student at the Indiana University Bloomington, pursuing a master's degree in Data Science. Additionally, I am currently employed as a Research Assistant, working under the mentorship of Professor Roni Khardon.
Email: santhoshpatil9798@gmail.com
Phone: (812)-778-4649
Python, R, C, SQL
PostgreSQL, MySQL, MongoDB, Neo4j
PyTorch, TensorFlow, Keras, NumPy, SciPy, Scikit-learn, Pandas, OpenCV, Matplotlib, Seaborn, NLTK, PySpark
Matplotlib, Seaborn, Tableau
AWS (EC2, IAM, S3, ELB, EMR, ECR, Lambda), Amazon SageMaker, Docker, Kubernetes
Git, CI/CD, FastAPI
Normalizing Flows, VAEs, GANs, LLMs
Linear & Logistic Regression, Decision Trees, Random Forest, SVM, Naive Bayes
Bayesian Inference, Variational Inference, Computer Vision, NLP
Refined and applied a Variational Probabilistic Matrix Factorization model for movie rating prediction, enhancing performance by 10% through strategic initialization and training of an EM Algorithm, with detailed assessment of RMSE improvements.
Improved topic modeling accuracy and posterior approximation by combining word and topic embeddings with a Variational Autoencoder and analyzed LDA vs. ETM models using Topic Perplexity and Diversity for superior topic extraction.
Engineered a cutting-edge deep learning model by integrating Error Level Analysis and ResNeXt, enabling highly accurate detection of image manipulation; outperformed industry benchmarks and established new standards in image forensics with 94% accuracy score.
Developed a model for the automatic detection of staff lines, musical notes, and rests within a sheet of music.
The assignment encompasses tasks such as handling transformations between the 3D world and 2D images, as well as the implementation of Markov random fields to infer depth from stereo images.
Developed a Tropical Cyclone forecasting model using a hierarchical encoder-decoder architecture, inspired by Earthformer model, for efficient high-dimensional data processing.
Crafted a hybrid model that combines Convolutional LSTM and Vision Transformer with cuboid attention blocks and global vectors, improving complex Spatiotemporal data management, resulting in precise predictions and a 35% reduction in test loss.
Analyzed model performance with RMSE and Focal loss, achieving an F1 score of 0.82.
Implemented an Ensemble model to enhance predictions with long lead times, achieving a 0.75 score at 42 hours.
Explored probabilistic Bayesian approaches for better uncertainty estimation and to simplify the training process compared to multiple ensemble models.
Improved decision-making by refining data on weather variables like wind speed and temperature at different sea levels, boosting forecasting accuracy by 16%.
Designed Privacy Center in Salesforce for customized data handling, complying with GDPR. Spearheaded an archiving system linking Privacy Center and Heroku Connect, speeding up data transfers to Heroku by 30% while ensuring security and compliance.
Established a stringent data retention policy for securely isolating production data, reducing the risk of unauthorized access and ensuring compliance with data protection regulations.
Investigated methods to enhance SLAM algorithms through Optical Flow, aiming to generate higher quality maps.
3D Instance and Semantic Segmentation: Modified techniques for 3D identification by applying 2D convolution to extract features from RGB images, succeeded by merging these features with geometric information in a 3D fusion process.
Enhanced the 3D map reconstruction in the Scan Complete study by integrating the Truncated Signed Distance Field method to patch gaps and eliminate blockages, achieving a refined, high-resolution map that boosts robot-environment engagement.
Expanded the capabilities of the PolygonRNN Auto - Annotation model by incorporating the YOLO Algorithm for precise object recognition and classification, accelerating annotation speed by 50% and elevating overall annotation quality.