VASU GUPTA

ML for Diabetes:

• A project for the AI/ML club at ASU where we used a naive Bayes implementation with Bayesian networks to check if a user has type 1 diabetes based on blood data
• Accomplished auc of 0.68 within 2 months and 0.705 within 4 months which is within 94th percentile for such models
• Built and trained using ASU clinic and Kaggle datasets
• Used matplotlib, scikit-learn, pandas and NumPy to build a naive bayes implementation from scratch.

Problem & Motivation
Standard RNNs (LSTMs/GRUs) allocate separate bias parameters for each gate and hidden unit, leading to over-parameterization and overfitting—especially in low-data or resource-constrained settings.

Innovative Approach
• Gate-level & layer-level grouping: share bias vectors across all input, forget, output, or per-layer units rather than per unit.
• Learned clustering: introduce a regularization term that automatically ties similar biases into a small number of groups.
• Soft vs. hard tying: explore both gradual (soft) and strict (hard) parameter sharing during backpropagation.

Key Contributions
• Compact RNNs: achieve substantial reductions in bias-parameter count with minimal impact on perplexity (language modeling) or MSE (time-series forecasting).
• Generalization Gains: improved robustness under limited training data by avoiding overfitting.
• Open-Source Toolkit: PyTorch and TensorFlow implementations, complete with scripts for automatic grouping and benchmarking.

Impact & Applications
Enables deployment of high-performance sequence models on edge devices and in any scenario where model size and generalization are critical.

Built a website for CloudTenX

• Built an automated CI/CD Pipeline with GitHub, Vercel, and Checkly
• High quality code build, test and deploy within 5 minutes of pull request merge
• Used NodeJs, and react to build the webapp, also contributed to UI/UX design and implementation

• A project with a goal to automatically build an API given a schema
• Implemented a Multi-Tenant database allowing for the creation and removal of Tenants in an SQL database
• Each Tenant had a record database keeping track of their posted schema and resulting API endpoint generated
• Used React and node to create an API that takes a users schema and creates and returns an API endpoint based on the schema