This page brings together my current machine-learning research, open-source computer-vision work, and production AI systems. The projects span probabilistic modeling, generative AI, causal inference, NLP, recommendation, safety, anomaly detection, and reproducible ML workflows.
Research projects emphasize clear modeling assumptions and repeatable evaluation; industry projects emphasize measurable product impact, reliable deployment, and collaboration across engineering and operations.
Developed probabilistic methods for improving fairness in pre-trained image generators without retraining the large base model from scratch. The work led to a SafeGenAI Workshop publication at NeurIPS 2024.
Develop tractable probabilistic-modeling methods for individual treatment-effect estimation and counterfactual inference, with attention to reproducible evaluation and clearly stated assumptions.
Partner with the Arizona Department of Emergency and Military Affairs on predictive modeling for natural-disaster forecasting, including reproducible pipelines, model evaluation, and technical documentation.
Built and published an open-source residual-diffusion deep-learning framework for image watermarking, connecting the research method to a codebase that others can run and extend.
Combined convolutional networks, semantic segmentation, saliency estimation, and content-aware resizing so retargeted images preserve visually important regions.
Contributed to deep-learning brain-tumor segmentation and to a method for lossless angiogram foreground compression while preserving the visual quality of the background.
Worked on instance-aware artistic image filtering with convolutional networks and on seam-carving methods that control the positional distribution of removed seams.
Python
PyTorch
TensorFlow
C
C++
Linux
Git
Docker
MATLAB / OctaveDesigned, trained, shadow-tested, and deployed an automated advertisement-acceptance system across seven product verticals. It reached 25% automation and saved work equivalent to 50 operations employees.
Developed chat-safety models from text and numerical features. The fraud model improved precision by 8%, while the harassment model improved precision from 0.26 to 0.44.
Built an anomaly detector using text embeddings and time-series behavior to identify abnormal advertisements and support operational and business decisions.
Designed collaborative-filtering and content-based recommenders, evaluated them through A/B tests, and increased job-seeker engagement by 11%.
Built production workflows with Python, PySpark, and Airflow and evaluated FAISS, Milvus, and Elasticsearch for large-scale recommendation and vector-retrieval use cases.