Natural Language Interfacing w/ Cozmo using GPT-3.5

Built a Python-based system that takes user text input and automatically generates a syntactically-correct Finite State Machine (FSM) file using GPT-3.5, and executes the program on the robot in real-time
Optimized accuracy and program validity through prompt engineering to base model gpt-3.5-turbo
Found that GPT occasionally showcased its own flair - when prompted to “spin in a circle”, it generated the following:

robot

Exploring Golf Analytics From Trackman System: Consistency and Clustering Analysis

Assisted the CMU varsity golf team (Div III) as a statistical consultant for a capstone project
Collaborated in a small team to analyze indoor golf simulator data from the state-of-the-art TrackMan system to provide actionable insights to the golf team coach, for the individual athlete and overall team
Wrote code in R (using packages such as ggplot2 and the tidyverse suite) to produce high-quality data visualization to complement in-depth statistical modeling & analysis
Our poster won second place at CMU Meeting of the Minds, a showcase competition for undergraduate research

driver spin_loft_club

The first project of my club (Tartan Neurotech) in the NeuroTechX 2022 Student Clubs Hackathon, an international neurotech competition
Led student teams to design & implement hands-free painting software, using ML to decode SSVEP brain signals and increase accessibility for motor-impaired users, with the goal of integrating eye-tracking control with neural signals for a seamless experience
Built a basic software in Python to test SSVEP signals with rotating icons, recorded signals with OpenBCI Ganglion hardware (below):

plot_from_flashing_15hz

Built an end-to-end speech recognition system using attention mechanisms (from the Listen, Attend and Spell paper), implementing a transformer architecture using NumPy and PyTorch to enable high-accuracy speech transcription on the LibriSpeech dataset
Acheived a final (private score) Levenshtein distance of 12.23 on Kaggle, after training for 30 epochs and employing techniques such as label smoothing, frequency masking, locked dropout, and weight decay
Part of the final assignment for Introduction to Deep Learning (11-485) @ CMU

LAS_1 LAS_2 LAS_3

Researched and implemented the Drift-Diffusion Model (DDM) of cognitive psychology (a model for decision making in the brain) using the Sequential Probability Ratio Test to compare hypotheses
Extended the DDM by generalizing the model to remove underlying assumptions (distribution, mean, variance, etc), and applied supervised machine learning using a Support Vector Machine (from scikit-learn) to enable the model to learn from data

DDM DDM_2

Implemented the Expectation-Maximization algorithm for Gaussian Mixture Models (GMM) in MATLAB to perform neural spike train sorting, as part of an assignment for Neural Signal Processing (42-632) @ CMU
Built a unsupervised learning system for neural decoding (classification of spike trains), reached rapid convergence on clusters for real neural data provided by the Shenoy lab

NSP_1 NSP_2

Investigated a Gabor-filter linear regression approach towards decoding images from highly sparse 2-photon calcium imaging data from the primate V1 cortex. Final results were mixed with a correlation coefficient of 0.43 between original and reconstruction, but worked well for simple patterns such as a leopard’s fur (see below)
Trained model on an existing dataset gathered in 2018 by Dr. Shiming Tang et. al. (Large-scale two-photon imaging revealed super-sparse population codes in the V1 superficial layer of awake monkeys), includes gigabytes of neural responses to natural & artificial visual stimuli
Conducted research as part of a summer role in the Lee Lab for Biological & Machine Intelligence

gabor_wavelet tuningCurves10

original reconstruction gaborReconstruction

Developed software for an NSF-sponsored virtual chemistry simulator (ChemCollective) which allows students and educators to design and perform their own lab experiments. Part of a collaborative project between 15+ universities, and free to use for teachers around the world
Leveraged skills in JavaScript and graphic design to expand the inventory of digital chemistry equipment, adding both form and function

vlab_3_50

Developed an educational interactive protein-folding game using Python to teach the users about the interactions between amino acids in the primary structure of proteins. The game presents each peptide as a challenge, to fold in accordance with biological properties
Animated using Tkinter, a Python graphics library. Implemented a basic “autosolver” using backtracking to serve as a reference for students
Part of the final term project for Fundamentals of Programming (15-112) @ CMU

maxresdefault_50