About Me

Senior Research Scientist at Netflix (New York).

Machine learning research scientist specializing in generative AI, LLM reasoning, uncertainty quantification for deep models, and sequential decision-making. At Netflix, Spotify, Princeton, and Columbia, I have advanced techniques for reasoning, structured event modeling, calibration and Bayesian uncertainty for large networks, offline reinforcement learning, and exploration. My work appears in NeurIPS, ICML, AISTATS, Machine Learning Journal, WWW, RecSys, and KDD, and has been deployed to systems serving hundreds of millions of users.

Previously: 

• Senior Research Scientist, Spotify, New York

• Adjunct Professor, Columbia University

• Postdoctoral researcher, Columbia University and Princeton University

• PhD in machine learning for spatiotemporal modeling, University of Southampton

• MSc (Artificial Intelligence), Imperial College London

• BA (Computer Science), Oxford University

News and updates

Research Interests

Deep Uncertainty Quantification

Uncertainty is an inherent aspect of machine learning models and is crucial to quantify for downstream insights and decision-making. As we scale up deep learning both in number of parameters and data size, the classical statistical methods become intractable or less relevant. I am particularly interested in rigorous methods to estimate epistemic uncertainty in deep learning.

causal recommendation

In recent years, the causal challenges present in recommender systems have been increasingly recognized. My work looks at how to train models with data that are confounded by the recommender and other biases.

Variational inference

Variational inference and variational autoencoders are widely used methods for performing approximate Bayesian inference on large data sets. My research is about how to perform inference on streaming data and how to deal with the non-convexity of the variational objective.

spatio-temporal probabilistic modeling

Spatio-temporal data require new models and decision-making techniques to deal with non-exchangeability. My research proposes methods in the areas of anomaly detection, reinforcement learning, and overcoming data sparsity in temporal patterns.

The d-speculation blog

AbouT

I maintain a blog called D-Speculation about ML, statistics, and research. I also write about Anything Else.