I’m a graduate researcher in Data Science at Fordham University, working at the intersection of generative AI, 3D understanding, and educational technology. I build end-to-end machine learning systems that go from prototype to publication — with a focus on multimodal models, diffusion learning, and low-resource inference.
My work spans open-source implementations, real-world product pipelines, and peer-reviewed research. I'm currently exploring ways to make AI more interpretable, controllable, and practically useful — especially in the context of education and healthcare.
-
Local Prompt Adaptation for Style-Controllable Diffusion
Proposed Local Prompt Adaptation (LPA), a training-free method for injecting content and style tokens in diffusion U-Nets to improve layout and stylistic control. -
RectifiedHR: High-Resolution Diffusion via Energy Profiling and Adaptive Guidance Schedule
Developed adaptive CFG schedules to stabilize diffusion sampling, improving image sharpness and reducing guidance artifacts. -
Limitations of NERF with pre-trained Vision Features for Few-Shot 3D Reconstruction
Conducted a systematic evaluation of DINO-enhanced NeRF models, revealing that pre-trained features de-grade performance in few-shot 3D reconstruction.
Languages: Python, SQL, Bash
Frameworks: PyTorch, TensorFlow, Scikit-learn, LangChain, Flask, FastAPI
Tools: OpenAI API, Play.ht, FFmpeg, Weights & Biases, Docker, MongoDB
ML Domains: Diffusion Models, LLMs, CNNs, Vision Transformers, MLOps, Explainable AI
Deployment: GitHub Actions, Streamlit, Gradio, Flask APIs
Reimplemented Oxford's NoPropDT model using layer-wise local denoising blocks (no backprop) and achieved 99% test accuracy on MNIST. Modular PyTorch pipeline supports sample reuse, cosine schedules, and extension to CIFAR.
Developed a text-to-video generation system using OpenAI (script), Play.ht (voice), and FFmpeg (rendering), with dual pipelines for PDF lectures and prompt-based generation. Integrated RAG-powered chatbots and user-facing professor/student apps.
Benchmarked VGG, Inception, and DenseNet models on 7-class skin lesion classification task, achieving 83.9% test accuracy with an ensemble-based pipeline and detailed EDA/metric analysis.
I'm open to research collaborations,full time roles, internships, and applied AI projects. Feel free to connect!