Hi! I'm Michael.
I'm interested in applied math, robotics, and neuro. Currently, I'm researching ways to integrate large language models and computer vision techniques to create generalist robotics. I'm also a member of the perception and vehicle dynamics/control subteams of the Cavalier Autonomous Racing team at UVA.
In the past, I did bioinformatics research at Dartmouth and game theory research at the US Army Research Lab. I also contracted with Kyron Learning to develop large language model math tutors.
I like running and used to do a lot of speedcubing.
Robotics Research Assistant, University of Virginia
Dec, 2023 - May, 20262 years 6 months
Student researcher investigating capability for robots to follow natural language instructions and perform in-context learning. Implemented imitation learning algorithms to solve tasks in RLBench (using PyTorch). Used ResNet-18 and CLIP as visual encoders.
Autonomous Racing Team, University of Virginia
Sep, 2023 - May, 20262 years 8 months
Developing software to race against opponent teams in the Indy Autonomous Challenge.
Implemented model predictive control in C++ through differential equations representing vehicle model. Improved final displacement error by 40% with gradient-based error minimization of model parameters against measured data.
Wrote Python code to perform uniform cost search over future trajectories to optimize car speed while maintaining safe distance from opponent vehicle.
Software Engineering Intern, Kyron Learning
Dec, 2022 - Oct, 202310 months
Creating language models to serve as effective tutors for math and science.
Write Python API for large language model-based tutoring system to supplement at-home learning. Improved factual and quantitative accuracy through semantic similarity search in vector database and Wolfram Mathematica integration.
Created React website for lectures with interactive AI-teacher checkpoints. Used automatic prompt optimization approach to generate better practice questions than the baseline (70% of results preferred to original).
Reinforcement Learning Researcher, U.S. Army DEVCOM Army Research Laboratory
May, 2023 - Aug, 20233 months
Student researcher using reinforcement learning to train robots acting in uncertain, adversarial scenarios.
Used PyTorch to train graph neural networks with reinforcement learning for robotic path planning against an adversary. Created 97% efficiency gain compared to traditional approach
Poster presentation at AAMAS 2024 (International Conference on Autonomous Agents and Multiagent Systems).
Computer Vision Researcher, Dartmouth-Hitchcock
Jun, 2022 - Aug, 20231 year 2 months
Student researcher using computer vision techniques in colorectal cancer research.
Primary author of 3 journal articles to predict cancer-related cell activity based on images using convolutional and graph neural networks.
Trained cell detection models (convolutional neural networks: Mask R-CNN and Faster R-CNN) for histology images with 80% mAP.
Leveraged GPU cluster to train machine learning models with Slurm.
Visualized results with Seaborn, Matplotlib, and Pandas. Interpreted results with dimensionality reduction and clustering: UMAP, t-SNE, and HDBSCAN.
Embedded Software Developer, Solar Car Team at UVA
Aug, 2022 - May, 20239 months
Writing embedded software to run on a solar-powered car.
Developed a regenerative braking system to recycle 9% of energy and a PID controller to stabilize driving
Undergraduate Research Assistant, University of Virginia
Oct, 2022 - Feb, 20234 months
Wrote C++/CUDA code to accelerate simulation of atomic forces and approximate them with neural networks.
Implemented multilayer perceptron with backpropagation in CUDA and contributed to popular open-source repository: GPUMD
Nov, 2022 - Dec, 20221 month
In this project, I sought to create a way to use language models for research-oriented writing. It provided the following primary features: (1) A writing assistant with drag-and-drop text completions from external documents, with an integrated Bing search tab that enabled quick access, and (2) A research assistant that enabled a natural language summary of top Bing search results with citations. This was implemented as a React app and a set of AWS Lambda functions. The backend used OAuth2 for authentication and a MongoDB database to store user profiles. I used OpenAI's (now-deprecated) text-davinci-003 model for text generation. I eventually reached 10 paid users (yay!), and I shut down the project when ChatGPT was released.
Jun, 2021 - Nov, 20215 months
An app to simplify carpooling: students can easily find people to carpool with for after-school events, eventually getting 300+ users. Worked with school administration and sports coaches to get schoolwide adoption. Helps students who live far away from a school to have the same access to after-school activities as people who live nearby, especially if they are unable to drive themselves. Stack used: Frontend: Typescript React Backend: Node.js/Express, Prisma ORM, PostgreSQL database Hosting: DigitalOcean droplets, VPC for database security.
Demo: https://www.youtube.com/watch?v=wJrC0Eozu6Y
Frontend Github: https://github.com/myfatemi04/wheelshare-frontend
Backend Github: https://github.com/myfatemi04/wheelshare-altbackend
Aug, 2021 - May, 20229 months
Created a Raspberry Pi device to locate sounds within 8º mean absolute error. Uses custom sound localization algorithm implemented in C that filters sounds by frequency, compares the phase of each signal, and triangulates an expected 3D position.
Mar, 2020 - Mar, 20202 days
An app that enables the people with disabilities to control their computer mouse with their eyes: after a quick calibration step, the computer would detect which section of the screen they were looking at with 90% accuracy and move the mouse there. Then, via voice recognition provided by Google Cloud, they could click on those parts of the screen, type, search the web, and perform other tasks. Won 1st Place at HooHacks 2020, a hackathon organized by students at the University of Virginia. Utilized Haar cascades, DLib facial recognition, a custom iris detection algorithm made with OpenCV, and voice recognition from Google Cloud.
Video: https://www.youtube.com/watch?v=PvBQVCoy1MQ
GitHub: github.com/myfatemi04/EyeOS
Computer Science, Bachelor’s of Science, University of Virginia
May, 2022 - May, 2026
Deep Learning
Natural Language Processing
Advanced Algorithms
Computer Systems
Probability
Cognitive Neuroscience
Optimization
Partial Differential Equations
Complex Variables
Real Analysis
Linear Algebra
Computer Science, Computer Science, Thomas Jefferson High School for Science and Technology
Dec, 2018 - Dec, 2022
Artificial Intelligence
Computer Vision
University of Virginia Entrepreneurship Cup: Winner , University of Virginia
Awarded on: Oct 01, 2022
Won $1000 for pitching an AI tool to help people prepare for interviews.
MIT BattleCode - 5th Place High School Team , MIT
Awarded on: Jan 31, 2022
Competed in MIT's BattleCode competition
https://github.com/2022tgoel/Battlecode2022
University of Virginia Hackathon: 1st place out of 100 teams , HooHacks 2020
Awarded on: Mar 31, 2020
Made a program that controlled a computer with eye movements and voice commands.
Demo: https://www.youtube.com/watch?v=PvBQVCoy1MQ
2x AIME Qualifier , Mathematical Association of America
Awarded on: Jan 01, 2021
Qualified for the AIME twice - top 2.5% of students in the AMC 10, top 5% of students in the AMC 12
Mentor, Web Developer, Director of Events, Codefy
Apr, 2020 - Apr, 2021
Organized workshops in computer vision and game development, mentored students in Python programming, taught web development class, created an enrollment system for 2000 students and 200 mentors
Deceptive Path Planning via Reinforcement Learning with Graph Neural Networks., International Conference for Autonomous and Multi-Agent Systems
Published on: May 06, 2024
We create a reinforcement learning algorithm to make robotic path plans in adversarial, uncertain environments.
Inferring Spatially-Resolved Transcriptomics from Cell Detections , Under review
Published on: Aug 01, 2023
We explore the explainability and accuracy of using cell detections as intermediate features when making predictions about tissue samples.
The Overlooked Role of Specimen Preparation in Bolstering Deep Learning-Enhanced Spatial Transcriptomics Workflows , Under review
Published on: Aug 01, 2023
We explore the ability for improved slide staining techniques (like automated staining) to improve the ability for machine learning models to make inferences about tissue samples.
Spatial Omics Driven Crossmodal Pretraining Applied to Graph-based Deep Learning for Cancer Pathology Analysis , Pacific Symposium on Biocomputing
Published on: Jul 01, 2023
We explore the utility of leveraging spatial transcriptomics data with a contrastive crossmodal pretraining mechanism for cancer staging, lymph node metastasis prediction,
survival prediction, and tissue clustering analysis.
Inferring spatial transcriptomics markers from whole slide images to characterize metastasis-related spatial heterogeneity of colorectal tumors: A pilot study. , Journal of Pathology Informatics
Published on: Jan 01, 2023
We assess the ability for machine learning models to infer gene expression levels across tissue samples. We prototyped and compared several convolutional, transformer, and graph convolutional neural networks to predict spatial RNA patterns at the Visium spots under the hypothesis that the transformer- and graph-based approaches better capture relevant spatial tissue architecture. We further analyzed the model’s ability to recapitulate spatial autocorrelation statistics using SPARK and SpatialDE.