Portfolio

Machine Learning

Notable achievements (all in TensorFlow):

Built and trained a transformer to translate Portuguese sentences to English with a bleu score of 0.28.

Created Resnet18 from scratch utilizing residual connections between convolutional layers on MNIST.

Also created a U-Net from scratch with the purpose of pixel-wise segmentation for vein detection in iris scans.

Worked on an LSTM-autoencoder for anomaly detection in time-series sales data on my co-op during Summer 2023.

Notable achievements (most in Numpy):

Built an autoencoder and feed forward neural network with custom made neural network architecture.

Designed a single perceptron network to detect images of 0 or 1.

Programmed machine learning algorithms from scratch such as minimax for tic tac toe.

Built decision trees and random forest models from scratch.

I am performing research on creating a simulated robot that takes in multimodal data and attenpts to find a target. This research is based on subterranean exploration in unknown environments. My team is attempting to train the model initially using imitation learning.

Robotics

In the robotics team, I was in charge of updating our robot Dokalman's software from ROS1 to ROS2. This meant rewriting the whole codebase. I wrote code to talk to the GPS and motor controllers over USB via UART, conntected to camera feed for processing using OpenCV, as well as starting on Nav2 by simulating the robot with URDF/XACRO.

Additionally I rebuilt the whole network from scratch as the base station's network links to Dokalman's network. This allows for easy remote access to Dokalman as well as providing internet to it on the field. I also worked on some of the mechanical hardware design and electrical work throughout this process.

I worked with the UC Football Marketing department to take the T-Shirt cannon onto the field and shoot shirts to the student section. This cannon was a senior design project, so I did not build it, but I improved many of the electrical aspects of the robot.

I learned how to take a live camera feed and apply a transformation to the image so it is top-down instead of angled. I also learned how to detect lines from images using OpenCV. I placed these processes into C++ ROS2 nodes and created a reusable vision pipeline for our robot to utilize.

The team worked on a sub-project which is this RC Car. We set it up to chase barrels of specific colors. It is also driveable over ROS2 using an offboard computer with a USB controller plugged in on the same wifi network.

Other Cool Achievements

Award in 2019 and received from Troop 123 located in Doylestown OH. Completed an Eagle Scout project that required over 500 man hours and over $2,500 of funding as well as many material donations to complete the project. The project finished in 2019 and is still operational.

I dabbled with KiCad to make custom coasters for my house. It was a fun project and even though the boards don't work I learned a lot!

Awards received at the National Robotics Competition. One group from each competition type is nominated for innovative robot design and programming compared to teams in the same competition. The competition is judged by Honda Engineers. Awarded for Bot Ball (2017), Maze Solving (known layout - 2018), and Micro Mouse challenges (2019).

I built this website from sctratch using HTML/CSS/Javascript. I've also pulled the image carousel from the bootstrap examples. This site is hosted on GoDaddy and I push/pull changes using Git.

I built the CNN in the AI Demo page from scratch using TensorFlow. I trained it on the MNIST handwritten digits dataset and saved the model to a web-compatible format with TensorFlowJS. The model parameters are saved on my website and is making the predictions you see on the page. Predictions may not be 100% accurate every time.