My Projects

CasaLLM: An LLM built from scratch.

I built CasaLLM, a 350M parameter LLM built from basic PyTorch elements. It underwent pretraining, fine tuning, and RLHF over the course of 70 hours, and includes implementations of RoPE embeddings and kv caching.
See Full Blogpost Live Demo

Custom CLIP

I designed and trained a custom implementation of CLIP, which serves as the foundation for many of today’s multimodal systems.

This was a 34 million parameter model coded from scratch using PyTorch and trained on 3 million image-caption pairs. An RNN was used as the text encoder to explore its performance in the CLIP architecture.

See the full blogpost.

Master’s Thesis: Contactless Current and Voltage Detector

For my Master’s thesis at MIT, I developed a contactless detector that used signal processing and machine learning to infer current and voltage levels in a set of cables.

The goal was to replace expensive Hall Effect sensors with inexpensive magnetic field point measurements, using algorithmic methods to remove interference from external magnetic field sources.

Methods applied included Autoencoder Neural Networks, Linear Regression, Polynomial Regression, non-linear solvers, and Generalized Least Squares.

See the full thesis.
US Patent No. 12085591

DropAlan: A Novel Dropout Variant

For my 6.860 (Statistical Learning) final project, I developed a new dropout variant called DropAlan.

Instead of pairing traditional activation functions with standard dropout, DropAlan directly assigns each neuron a dropout probability based on its input values. In expectation, this mimics the behavior of ReLU activations while avoiding issues like dead ReLUs and making the method easier to analyze theoretically.

Tested on MNIST and CIFAR-10, it achieved similar accuracy to traditional dropout while converging as much as 20% faster.

See the full report.

OCR System

For my 6.869 (Computer Vision) final project, I designed an Optical Character Recognition system. I worked with a partner, Andrew Montanez.

In this project, we trained a VGG-16 variant CNN to classify letters in an image. Candidate image sections were isolated for classification using edge-detection techniques with the help of the OpenCV library.

Pole-Climbing Robot

As part of the MIT Class 2.007 (Design and Manufacturing), I designed and built a remote-controlled robot to perform various tasks in a field, including pulling levers and moving sand bags.


The robot was built out of aluminum plates and I used several machine shop tools, including a mill and lathe, to construct the robot. I was also required to perform physical analysis of my robot design and justify my design decisions.

MASLAB Robot

I competed in the January 2015 MIT MASLAB Robotics competition. With two teammates, I built an autonomous robot designed to participate in a competition which involved picking up red and green blocks from a playing field and depositing them in a target area.


The robot was built using MDF and laser cutting, and was designed in Solidworks. We used the OpenCV library to process a webcam feed to enable the robot to autonomously locate and pick up the blocks on the field.

Let’s Connect