LLM Evaluation

2025

This project consists in exploring the ways to evaluate machine outputs. We explore lexical overlap metrics, embedding based metrics as well as LLM-as-a-judge as ways to evaluate a machine output.

Audio Feedback Generator

2025

This project is an AI-powered audio feedback generator. It analyzes the content as well as the delivery of the recorded audio. The system operates through a cloud pipeline: a React frontend handles audio recording and uploads, which are then processed by a series of Cloud Run microservices. The backend converts and cleans the audio, performs signal analysis and speech-to-text transcription, and then uses a Gemini LLM to generate detailed feedback. All communication and real-time status updates for the user are managed securely through Google Cloud services, including Pub/Sub and Firestore.

The NBA's 3-point Revolution visualized

2025

A visual deep dive into how shot selection in the NBA has evolved over 20 years—from Mid-Range dominance to perimeter precision.

Reinforcement Learning - Maze

2022

A Reinforcement Learning agent was trained to navigate a maze environment—from the top-left corner to the bottom-right—while avoiding obstacles. Several RL algorithms were explored and compared to optimize pathfinding performance in the presence of walls and barriers.

A Historical Analysis of Harmonic Progressions using Chord Embeddings

2021

This project explores chord patterns in classical music by applying Word2Vec and LSTM models to a large historical dataset, revealing stylistic differences between composers and evolving harmonic trends over time.

Finger Tracking and Digit Recognition

2021

Tracking was done using OpenCV's cv2.matchTemplate function, which applies template matching—most effective with plain black or white backgrounds. For digit recognition, a Convolutional Neural Network (CNN) was trained using PyTorch.