Click Below to Try Out My Projects
AI Insurance Agent (WIP)
Conversational agent to explain and quote health insurance.
RapGPT2.0
LLM that generates rap lyrics & audio in Eminem's style and voice.
Retinal Disease Classifier
CNN model that classifies retinal images for disease diagnosis.
Dive Deeper into the Details of Each Project Here!
AI Insurance Agent (Learn More...)
An interactive assistant designed to help users navigate and select health insurance plans.
Built with LangGraph, this agent leverages a combination of retrieval-augmented generation (RAG), function-calling, and >decision-oriented dialogue flows.
RapGPT2.0-Eminem (Learn More...)
Built on GPT-2 with 124M parameters, this model was trained on AWS EC2 with LoRA fine-tuning to generate Eminem-style rap lyrics.
It includes KV-caching, flash attention, and quantization for fast inference.
Evaluated model outputs with a BERT-based classifier to ensure stylistic fidelity with Eminem Lyrics
Integrated with ElevenLabs for audio output and deployed with a custom UI for interactive lyric generation...
Retinal Disease Classifier (Learn More...)
Developed a hierarchical classification model that identifies retinal diseases from medical images, achieving 85% recall in disease identification and 82% accuracy in classifying specific conditions.
The project involved building and training models using EfficientNet and Vision Transformers and was evaluated with the help of TensorBoard.
To optimize the model for deployment, model's size was reduced while maintaining high performance, achieving an average inference time of 0.5 seconds.
The final model was deployed on Hugging Face Spaces, allowing users to interact with the classifier via a simple, intuitive interface built with Gradio...
Other Projects
From Sratch
This repo includes from-scratch implementations of various machine learning frameworks, such as LoRA, GPT, Flash Attention, Neural Networks, Vision Transformers, EfficientNet, and more.
rapGPT 1.0
Trained a small decoder-only transformer model using with a custom tokenizer, with approximately 1 million parameters. The model performs very poorly which led to the rapGPT 2.0.
GF is Hungry-Fremont Edition
Developed a random restaurant chooser for restaurants in Fremont, where users can send a prompt. The model leverages FAISS to search a vector database and return restaurant matches based on the given prompt.
Poker Simulator
Created a poker simulator that calculates the win probability based on the cards in hand, the number of players, and varying community card combinations.
This Website
Built this website using Next.js,