Hi, I'm Morgan
Contact MeAbout Me
My introductionHi, I'm Morgan. I'm a software engineer at Meta, where I optimize machine learning models for their video recommendation system, enhancing user engagement and content relevance.
I'm passionate about AI, especially building it responsibly. I care deeply about AI safety, reliability, and the choices that go into deploying models at scale. I graduated from Brown University with a concurrent Bachelor's and Master's degree in Computer Science, and I've explored these interests through work at Amazon and IBM, as well as freelance work.
Outside of engineering, I'm a former ballerina and still love staying active in any form — Pilates, running, or just being on the move. I'm always down to wander into a new coffee shop, spend an afternoon shopping, or try something new with friends or family. These parts of my life keep me grounded and curious, and they shape how I show up as both an engineer and a teammate.
Projects & Research
PapersUtility of LLMs as The Foundation for Generalized RL Agents
Ollama • Llama3.2 • MiniGrid • Reinforcement Learning
Explored the feasibility of using out-of-the-box Large Language Models as policy generators for reinforcement learning agents, without domain-specific training. Developed LLM-RL, a novel framework with structured prompting mechanisms, memory components, and reflection-based feedback loops to enable text-based RL in MiniGrid environments.
Evaluated performance across Empty, Go-to-Object, MemoryS7, and DoorKey environments, comparing against DQN and PPO baselines. While LLM-RL demonstrated extreme sample efficiency in simple navigation tasks, it struggled with sequential planning and object interaction. Results highlight current limitations of pretrained LLMs as direct policy generators, underscoring the need for enhanced grounding mechanisms and task-specific training for effective RL decision-making.
Real-Time Yoga Pose Classifier
TensorFlow • PyTorch • OpenCV • Computer Vision
Developed and benchmarked multiple deep learning architectures for real-time yoga pose classification, including a custom CNN, EfficientNet-V2, ResNet-18, and YOLOv11. Built to address the challenge of automated form correction in virtual yoga sessions, which have seen unprecedented growth since 2020.
Trained models on a 5-class dataset using transfer learning with pretrained ImageNet weights, achieving significant performance improvements over the custom CNN. EfficientNet-V2 and YOLOv11 demonstrated the best performance. Deployed a complete real-time video pipeline with confidence thresholding and on-screen pose classification.
Dimensionality Reduction in Sparse scATAC-seq
Scanpy • PyTorch • Single-Cell Genomics
Evaluated the robustness of three dimensionality reduction methods (LSI, cisTopic, and scBasset) for single-cell ATAC-seq data analysis. Using a 69,248-cell human bone marrow atlas with 22 distinct cell types, we systematically tested performance under increasing sparsity levels (10% to 95% dropout) and synthetic batch effects.
Assessed methods using K-means clustering with metrics including Adjusted Rand Index (ARI), Adjusted Mutual Information (AMI), and Homogeneity Score. Results demonstrated that cisTopic consistently produced the highest-quality embeddings across all sparsity levels, maintaining superior clustering performance and developmental trajectory preservation compared to LSI and scBasset. This work provides a systematic framework for benchmarking dimensionality reduction techniques in sparse scATAC-seq datasets.
Scrabble Learning Model
OpenCV • TensorFlow • Deep Learning
Built an end-to-end CNN-based system for digitizing Scrabble board states from livestream video feeds, enabling real-time game analysis and strategic recommendations. The system employs custom image processing pipelines to handle varying lighting conditions, camera angles, and board configurations.
Trained character recognition models on the Chars74K dataset with custom augmentation strategies to simulate real-world Scrabble tile appearances. Used OpenCV for board detection, grid extraction, and tile segmentation. Achieved 97% character recognition accuracy with robust performance across different video qualities and board designs.
Zillow Zuperstars: Pandemic Housing Analysis
NumPy • Scikit-Learn • Data Science
Analyzed the pandemic's impact on residential real estate markets across the United States using Zillow's publicly available datasets spanning 2020-2022. Examined relationships between housing inventory, seller profit margins, and property price trends across approximately 900 cities.
Applied regression analysis and statistical hypothesis testing to identify correlations between inventory levels and profit margins. Used k-means clustering to discover four distinct market patterns: large cities with falling prices, small cities with rising prices, mid-sized cities with growth, and small cities with decline. Results provide insights into pandemic-driven migration patterns and their effect on housing markets, with findings validated through pairwise Z-tests confirming statistically significant cluster differences.