Applied Multimodal Team. A few things I've worked on:Veo: Latent diffusion model for video generation. I enabled high memory efficiency for large-scale deployment, and led efforts to deploy Veo (YouTube, Vertex AI and more, stay tuned!). I'm currently focused on enabling the fastest video generation experience with Veo.Gemini: Improved video understanding across millions of videos.Model Inference Optimizations: such as... added inference caching, reducing Gemini on video latencies by 90%. Developed memory-optimal upsampling on TPUs. Developed algorithms enabling low-resource latent decoding, decreasing memory footprint by over 80%.

Led R&D of privacy preserving inference and training of deep learning models across modalities.- Vision: object detection, face recognition, image retrieval, OCR.- Language: Generative text with LLMs, NLP fine-tuning tasks - Tabular: classification, xgboost, sensitive feature obfuscation.Formulated novel methods in stochastic representation learning, adversarial robustness, and multi-objective optimization. Core patents in submission.Lastly, architected the core Software Development Kit to enable integration of core technology and algorithms into any PyTorch based library, e.g. compatible with Hugging Face trainers, and academic research code.

Research on enabling secure inference and training for deep learning models across modalities -- image, video, text, tabular.

Engineering on the Responsible AI: Fairness team. Our goal is to provide tools and methodologies to address algorithmic bias in machine learning systems. Primary work on the team involves:- Developing core functionality of Fairness measurement tools, including statistical computation and scalability- Implementing privacy-first measurements, enabling large scale fairness measurements of ML models with respect to protected attributes.- Research Engineering to support applied research on Meta subsidiaries. - Proposing new measurements of fairness, particularly on high dimensional unsupervised tasks.

R&D on mitigating adversarial attacks on neural networks. Applied pruning and frequency domain analysis to study the relationship between robust/non-robust features. Supervised by Dr. Michael A. Warren.

Deep learning research on:- Explainability and bias of neural networks in Computer Vision. Advised by Dr. Qiang Liu. Additionally supervised by Dr. Ruth Fong at University of Oxford.Completed Thesis: Debiasing Convolutional Neural Networks via Meta Orthogonalization

Funded as a Graduate Teaching Assistant for the following courses:- Data Science Principles (Fall 2019)Responsibilities include writing and grading homeworks, holding office hours and grading exams.Topics taught include the theoretical basis for the following methods: Linear Regression, Decision Trees, Logistic Regression, SVM, Linear Discriminant Analysis, Naive Bayes, Boosting, Gaussian Mixture Models- Data Science Laboratory (Spring 2020)Responsibilities include grading homeworks and holding weekly labs (6 hours of instruction time/week).Topics taught in lab mainly cover practical considerations following the previous course, which include: Data preprocessing, Feature Engineering, Cross Validation, Data Sampling, Ensembling, Stacking. Also covered recent advancements in NLP and Computer Vision, introducing students to PyTorch.

Instagram Sharing Machine Learning team. - Created new Instagram Stories ranking models based off newly sourced labels- Implemented Lottery Ticket Hypothesis to apply neural network pruning to production ranking models- Tested different ranking methods for the Instagram Direct Reshare Share Sheet

E E 461P: Data Science PrinciplesResponsibilities:- Prepare programming homeworks in Jupyter notebooks- Hold office hours to cover key concepts- Weekly grading

Monetization Ranking team.Three main projects:-Created a new pooling layer for their deep ranking model using Caffe2. Significantly increased metrics and pushed to open source.-Develop new user side features to incorporate into feed ads ranking model.-Explored connections between ads side and user side features to merge and increase ranking metrics

CS 429H - Computer Organization and Architecture, Honors

At my time at ExxonMobil, I got assigned to work with Internal Audit. There, I worked with their Data Science team to build an anomaly detection system to be used in future audits. My main roles were to:• Write audit tests for feature engineering using Python and tabular models that analyzed invoices, work orders and purchase orders.• Develop the unsupervised anomaly detection model using Principal Component Analysis and Isolation Forest. • Create a new supervised model (using synthetic labels from unsupervised learning) that would support model improvement (as true labels come in) + continuous audits. In a Proof-of-Concept run, it had a 0.9 F-score with 100 anomalies out of a population of 120,000.I also supported two company side projects:• A document analysis tool that found similar lines between two documents that would allow engineers to build specification documents more quickly from Global Guideline formats.• A legal entity extractor that would find the ExxonMobil entity + Vendor in the preamble of a contract. This was a part of the summer Hack-a-thon and the project can be found here: https://github.com/kurtisdavid/LegalEntityExtraction

Under Dr. Cohen, MBChB, PhD at the UT School of Bioinformatics, I looked at how results of experimental cancer drugs in the lab can be shown to reflect the adverse drug effects reported to the FDA by drug manufacturers, health care providers, and clinical trials. I utilized data mining and web scraping to obtain the information needed, as well as machine learning algorithms to able to translate the correlations to a clinical setting.