01 / Experiment / 03.2025
Cat Diffusion
Model
A cat-image diffusion experiment built with PyTorch and a U-Net. Noise, structure, and the emergence of a recognizable form.
PyTorch / Diffusion / UNet
Research / Engineering / Experiments
Sound → Perception → Intelligence
Dahong Luo
01 / Selected builds
PROJECTS
Cat Diffusion
02 / Experiment / 03.2025
Quantum Neural
Network
A quantum–classical neural-network experiment combining PyTorch, Qiskit feature maps, variational circuits, and autoencoders.
PyTorch / Qiskit / ZZFeaturemap / RealAmplitude / Autoencoders
03 / Experiment / 03.2025
Fine-Tuning
SAM2
Fine-tuning the Segment Anything Model with PyTorch: an exploration of visual boundaries and image segmentation.
PyTorch / Segment Anything Model
04 / Experiment / 12.2024
Mathematical Figure
Transcriber VLM
A vision-language project for transcribing mathematical figures, using Llama and Unsloth on Google Cloud.
PyTorch / Unsloth / Llama / GCP
05 / Experiment / 10.2024
Embedded
Conditional GAN
An embedding-conditioned generative adversarial network, exploring how a learned condition can guide generation.
PyTorch / TCGAN / Embedding
06 / Experiment / 06.2024
Numpy Neural
Network
Neural-network and Transformer experiments in NumPy, with automatic differentiation as a building block.
NumPy / AutoGrad / Transformer
07 / Experiment / 05.2024
Cyberpunk 2077
Steam Review Analysis
An analysis of player reviews using language representations and machine-learning tools.
PyTorch / pandas / LightGBM / BERT
08 / Experiment / 01.2024
Reinforcement
Learning Pacman
A Pacman agent exploring reinforcement learning and Q-learning in a Pygame environment.
PyTorch / Reinforcement Learning / Q-Learning / PyGames
09 / IRONSIGHT Hackathon / 02.2026
AnchorVision
SLAM-based 3D object mapping from video
A 4D spatiotemporal mapping pipeline that combines visual SLAM and sparse UWB ranging to build persistent site memory without GPS.
ORB-SLAM3 / UWB / YOLO / LiDAR / C++ / ZeroMQ
- Integrated ORB-SLAM3 with sparse Ultra-Wideband ranging to maintain a globally consistent 3D site memory.
- Developed multi-agent alignment using UWB constraints as a global prior in a Maximum A Posteriori optimization solver, eliminating drift and enabling map fusion without visual loop closures or heavy overlap.
- Implemented 2D-to-3D semantic lifting by projecting YOLO detections through ray-casting and LiDAR depth-map fusion, creating a persistent object index keyed by time and coordinates.
- Built a C++ and ZeroMQ bridge to stream RGB-D and IMU data from iPhone LiDAR to a backend fusion service with sub-100 ms latency.
10 / HackUMBC / 10.2024
Grand Prize
Agent-Based
ELMS Copilot
A multi-agent educational assistant for syllabus parsing, course-content questions, and personalized quiz generation.
LangGraph / LangChain / Pinecone / PyTorch / Flask / React
- Engineered specialized agents with LangGraph and LangChain for syllabus parsing, content QA, and automated quiz generation.
- Architected a RAG pipeline using Pinecone vector databases for low-latency semantic search across heterogeneous course materials.
- Developed a custom question-generation engine using PyTorch and OpenAI LLMs, fine-tuned to extract pedagogical concepts from notes and generate tailored assessments.
- Improved retrieval using advanced chunking and metadata filtering to maintain context across complex subjects.
- Built a Flask backend and React/Vite frontend with Tailwind CSS, integrating PyTorch for pedagogical content generation and student analytics.
11 / Hacklytics / 02.2024
1st Place — Elevance Healthcare Challenge
Medical Data
Compilation System
A multilingual medical-document platform connecting document ingestion with natural-language database retrieval.
LLMs / SQL / OCR / Translation / Document Parsing
- Developed an upload pipeline that performs OCR, translation, and document parsing into SQL databases.
- Built a retrieval pipeline that converts natural-language queries into SQL, summarizes results, and suggests diagnoses.
- Created a web application for managing and analyzing medical documents in multiple languages and formats.
12 / HoyaHacks / 02.2024
1st Place — Best Digital Forensics Related Hack
Crime Event
Detector
An audio-driven video analysis system that finds critical acoustic events and turns long recordings into timestamped, searchable logs.
CNN / YAMNet-inspired Architecture / Audio Processing / MongoDB
- Developed a CNN-based audio segmentation model inspired by YAMNet to detect events such as gunshots, car crashes, and screams in lengthy videos.
- Curated custom datasets, applied data augmentation, and systematically tuned hyperparameters to distinguish target sounds from complex background noise.
- Built a pipeline for audio extraction, temporal segmentation, and timestamped event logs to reduce manual review in forensic and law-enforcement applications.
- Created a full-stack application with video upload, search, and MongoDB persistence for processed logs and detection results.
13 / Hack@CEWIT / 02.2024
1st Place — Spark of Genius Prize
Gesture Recognition
Control System
A multimodal interface connecting hand movement and voice to operating-system controls, a robotic claw, and an expressive air-guitar instrument.
OpenCV / Whisper / Hand Tracking / Robotics / Audio
- Engineered real-time OpenCV detection and tracking to map hand kinematics into operating-system commands for hands-free navigation.
- Integrated OpenAI Whisper for low-latency speech-to-text, allowing users to alternate between gestures and voice dictation.
- Built a hardware bridge from virtual spatial coordinates to physical actuation, piloting a custom robotic claw through real-time hand tracking.
- Developed continuous gesture-to-audio mapping based on hand positions, demonstrating responsiveness through an air-guitar interface.
02 / Research
RESEARCH
10/2025 — Present / Student Researcher
Neural Audio Codec
University of Maryland
Prof. Nirupam Roy · iCoSMoS Lab
A compact neural codec for real-time emergency voice transmission on low-power hardware.
- Developing a compact neural audio codec optimized for low-power microcontrollers, enabling real-time audio compression and reconstruction.
- Designing a low-power codec using residual finite scalar quantization and an EnCodec + Vocos-style model with < 1/20 the parameters of SOTA (~390K vs. 7M+), for emergency voice transmission in low-bandwidth, low-CPU settings: 2.3 kb/s and RTF 0.8 on a Raspberry Pi CPU.
- Created automatic training scripts, implemented Weights & Biases logging, and performed ablation studies.
06/2025 — Present / Research Assistant
Perception-Aware Audio
University of Rochester
Prof. Zhiyao Duan’s Group
Psychoacoustic compression and query-guided source separation for higher-fidelity neural audio.
- Integrated MP3/OGG psychoacoustic models to improve perceived quality in neural audio systems; analyzed LAME encoder stages and created Python bindings.
- Bound LAME MP3 and OGG Vorbis psychoacoustic features to Python through the PyTorch C++ interface, combining them with a finite scalar quantizer and Transformer-based flow matching for perception-aware, high-fidelity neural audio compression.
- Trained a U-Net-based joint audio separation model to distinguish multiple speakers and sources through natural-language queries. Led annotation and augmentation, iterated with PyTorch Lightning and Weights & Biases, and used grid search for hyperparameter tuning.
04/2025 — 09/2025 / Research Assistant
3D Human Reconstruction
University of Texas at Austin
Hezhen Hu, Postdoctoral Researcher · VITA Lab
Faster Gaussian-splatting reconstruction with measurable 3D mesh quality assessment.
- Enhanced PGSR-based 3D human reconstruction with a Gaussian Splatting-inspired loss and a faster testing pipeline. Profiled training bottlenecks and reduced training time by 80%.
- Introduced quantitative 3D mesh assessment using point clouds generated from output meshes and ground truth, automating validation and making output-quality tracking more reliable.
02/2025 — 09/2025 / Research Assistant
Research Writing Assistant
University of Maryland
Prof. Leo Zhicheng Liu’s Group
A retrieval-augmented workspace for literature review, scientific writing, and dataset exploration.
- Reviewed literature on generative-AI research assistance, collecting and analyzing papers to identify trends and key methodologies.
- Worked on a RAG-based writing assistant to help researchers write scientific papers.
- Contributed to a project helping event-sequence data scientists process datasets.
- Built metadata-driven graph re-ranking that improved retrieval accuracy, and integrated it into an interactive interface for faster evaluation.
09/2024 — 12/2024 / Student Researcher
Sound Event Detection
University of Maryland
LSTM Sound Event Detection Research · Honors Seminar
An efficient causal LSTM detector for real-time acoustic event recognition from waveforms.
- Conducted an Honors Seminar research project and wrote a paper as first author.
- Developed an LSTM-based sound event detection model that detects events from waveforms without relying on future data.
- Conducted an ablation study to better understand the model and improve it efficiently.
- Achieved real-time inference at 462 ms per 10-second clip with ~80% accuracy, compared with non-real-time CNN-LSTM state-of-the-art models.
- Later extended the system with a pretrained RAVE encoder and the Mamba temporal model, improving accuracy by 10% while preserving efficiency.
03 / Work experience
EXPERIENCE
Boston Consulting Group
San Francisco, CA
- Developed Unified Benchmarking Database, consolidating heterogeneous benchmarking datasets from across BCG into a standardized database and deploying an agentic retrieval layer for efficient access to benchmarking data.
- Onboarded 4 new benchmarking databases by building reusable ingestion scripts and an automated data-ingestion quality testing pipeline, while contributing to the agentic retrieval system and end-to-end data integration workflow.
University of Maryland, Engineering IT
College Park, MD
- Developed a Flask/React platform for university makerspaces to manage equipment access, enforce safety permissions, and track machine usage through secure authentication and streamlined workflows.
- Built a Gemini-based agentic code review system that retrieves relevant code via snippets and CLI tools, evaluates implementations against department policies, and aggregates outputs from three task-specific models with a fourth validation model, reducing review time by 30%+.
IZAI Corp
Tokyo, Japan
- Built and deployed a Flask/React RAG application on Azure using BM42 Hybrid Search and LlamaIndex; developed a metadata-driven agentic graph retrieval system with multi-reranking for efficient data retrieval.
- Fine-tuned MetaVoice TTS on ReazonSpeech with dataset cleaning and augmentation, and implemented 8-bit quantization optimizations that improved generation speed by 2×+ while maintaining audio quality
InfoDeliver Corp
Tokyo, Japan
- Developed an LLM-based agentic automation system using GPT-4 Vision to perceive UI elements and autonomously execute computer actions; built a real-time Whisper voice interface for natural-language task control and mid-task adjustments.
- Created and annotated a web UI component dataset to fine-tune YOLOv7, achieving 70% detection accuracy, and integrated the detector into a modular perception, chain-of-thought, and computer-action pipeline.
04 / Education & toolkit
EDUCATION
2023 → 2026 / Tokyo → Maryland → New York
-
—
University of Tokyo
Tokyo, Japan
B.S. studies in Natural Sciences · Subsequently transferred to the University of Maryland
-
—
University of Maryland
College Park, Maryland
B.S. · Computer Science major · Mathematics minor
GPA 3.90 / 4.0
-
—
Columbia University
New York
M.S. Electrical Engineering
Interests & skills
Interests
Audio Processing · Signal Processing · Generative AI · Large Language Models · Computer Vision · Natural Language Processing · Quantum Machine Learning
Languages
Python (7+ years) · C · C++ · Java · JavaScript · MIPS Assembly
Machine learning
PyTorch · TensorFlow · Keras · Qiskit · PennyLane · NumPy · scikit-learn · JAX · pandas · MATLAB · seaborn · Transformers · YOLO · Reinforcement Learning · Llama · LangChain · Unsloth
Technology stack
Azure · Docker · Git · Flask · REST APIs · React.js · Next.js · MongoDB · SQL · Hugging Face