Work

Novel research paper analyzing & quantifying LLM feature universality using SVCCA & RSA metrics and a sparse autoencoder dictionary-learning pipeline trained on Google Gemma 2-2B and Gemma 2-9B residual-stream activations. Unveiled striking results of mid-layer similarity and multi-token subspace similarity. Implemented feedback from Stanford and Berkeley Ph.D. reviewers and a Meta researcher. Presented findings at ACL 2025 in Vienna, Austria: the #1 ranked conference in the world for natural language processing/computational linguistics.

Enfora is a feature-rich accountability platform that enforces users' commitments with real financial stakes and evidence-based task validation. Users can define tasks with deadlines and financial consequences, upload image or document proof of completion, and receive AI-driven validation with confidence scoring, metadata checks, and an optional human review flow for disputed decisions. The platform is built with a React frontend deployed on Vercel, a Node.js (Express) backend deployed on Railway, and AWS services including DynamoDB with GSIs for efficient querying, S3 for secure evidence storage, Lambda functions for deadline enforcement and leaderboard maintenance, EventBridge Scheduler for automated task expiration, SES for emails, Redis-backed rate limiting, and Stripe for payment processing. The system features JWT-based authentication with refresh token rotation, role-aware routing, and scalable, fault-tolerant cloud infrastructure designed to handle adversarial usage patterns.

AI-powered Spotify playlist generator that creates and saves playlists based on user-provided natural language descriptions. Leveraging Python, Flask, the Spotify Web API, and the OpenAI API, the application searches for existing Spotify playlists that align with the user's preferences and generates a playlist of the Top-K most frequent tracks from such playlists using a heap and hashing techniques for optimized track ordering. Incorporates extensive error and edge-case handling for faulty inputs, API failures, and whitelist errors. Deployed on a simple web application developed with HTML, CSS, and JavaScript; hosted on Heroku.

Lung cancer risk classifier that utilizes a logistic regression model (Scikit-learn) trained on a custom dataset generated through a rule-based algorithm created using weighted features, biases, and mathematical relationships grounded in extensive research on the key contributors to lung cancer risk. Fine tuned the model using techniques like cross-validation, grid search, and gradient-based optimization to achieve 96.5% model classification accuracy. Deployed on a simple web application developed with HTML, CSS, and JavaScript; hosted on Heroku.

Java-based desktop keyboard macro application with a modern, intuitive GUI built using AWT and Swing. Vortex enables users to automate text input by specifying custom messages, repeat counts, and optional delays, offering a simple yet powerful tool for keyboard automation and productivity.

Developed computer vision models to assist autonomous vehicles using convolutional neural networks and transfer learning. Implemented a VGG16-based architecture trained on the CIFAR-10 dataset with TensorFlow and Keras, achieving 95% classification accuracy in object detection tasks.