Dahong Luo
Software Engineer & ML Enthusiast
Passionate about creating innovative solutions through code and machine learning
Passionate about creating innovative solutions through code and machine learning
I'm a Computer Science student at the University of Maryland with a passion for creative problem-solving and technology. My approach combines innovation with patience—I believe in connecting simple elements in innovative ways to develop unique solutions, whether through programming or collaborative projects. With experience in machine learning, software development, and a multilingual background, I'm dedicated to making meaningful contributions to the tech world.
My passion for creativity began in childhood with LEGO blocks, where I discovered that simple elements could be combined to create complex structures. This translated seamlessly to programming, where I view code as another medium for creative expression. Through hackathons and collaborative projects, I've learned that creativity is about connecting basic elements in innovative ways to develop unique solutions to real-world challenges.
Facing complex subjects like machine learning and mathematical modeling taught me that true understanding requires persistence. By methodically exploring background information, iterating through concepts, and seeking mentorship, I've developed the patience to approach problems thoroughly. This patience allows me to look beyond surface-level issues to address root causes, leading to more effective and detailed solutions in my work.
Honors Program, GPA of 3.81
Dean's List (Fall 2023 & Spring 2024)
Conducted research using LSTM for Sound Event Detection that doesn't rely on future data. Wrote paper as first author, implemented data processing pipeline, and conducted ablation studies.
Collected and cleaned a dataset of 788 mathematical images and corresponding transcriptions from the Mathverse dataset. Fine-tuned Llama 3.2 11B to improve the accuracy of transcribing mathematical figures, including complex handwritten notes and digital math figures.
Grand prize winner at hackUMBC 2024. Created a multi-agent RAG system with hallucination prevention and built a study suggestion model using PyTorch trained on synthetic data.
Recreated Generated Adversarial Network from the original paper. Improved the model by adding an Embedding Layer to generate images of arbitrary classes. Created a system that creates a seamless image transition video by gradually modifying the embedded values.
Created and trained UNet-based diffusion model to generate 32x32 images of cats. Collected and processed cat images from a Kaggle dataset. Trained the model on 30k images for 15 epochs, achieving 0.011 MSE loss with the test data.
Created Quantum Neural Network using Qiskit to perform classification tasks. Trained AutoEncoder to compress input data, enabling the quantum algorithm to capture more useful information. Constructed Quantum Neural Network using ZZFeaturemap and RealAmplitude with 8 qubits system.
Fine-tuned Segmentation Anything Model 2 by Meta to segment brain tumors given the CT scan of brains, achieving 0.64 validation IoU training for 12 epochs. Performed Exploratory Data Analysis using a dataset from Kaggle.
Built a PyTorch-inspired deep learning framework using only NumPy with custom layers, AutoGrad-like backpropagation, and full Transformer architecture, achieving 96.2% accuracy on MNIST.
Performed exploratory data analysis on steam review data from a Kaggle dataset. Trained DistilBERT model to predict the upvote/downvote from review text, achieving 94% accuracy. Trained Gradient Boosted Tree to regress the future upvotes of a comment with 2.956 MAE loss on test data.
1st place winner in Hacklytics 2024. Created a web app for managing medical documents with LLM-based pipelines for OCR, translation, and SQL database query handling.
1st Place of Spark of Genius Prize at Hack@CEWIT. Developed JarWiz, an innovative gesture and action recognition software enabling intuitive computer control through hand gestures and voice commands. Integrated Whisper model to support voice commands.
Designed an AI that autonomously learns to play Pacman using reinforcement learning. Engineered a fully functional Pacman game environment from scratch using Pygame with modular architecture for easy integration with AI agents. Developed a Deep Q Network (DQN) model with PyTorch.
Won "Best Digital Forensics Related Hack" at Hoya Hacks 2024. Developed a web app to pinpoint sound timestamps in videos for faster investigations. Built and trained a YAMNet-inspired audio classifier in PyTorch with a custom audio-to-image preprocessing pipeline.