I’m a Senior Scientist at Zscaler, working at the intersection of cybersecurity and AI. Before that I spent about a decade sitting at the overlap of networking, distributed systems, and machine learning, which is where my taste for composable, nearly-decomposable systems comes from. If you want the short version of what I do, it’s this: I try to take research ideas about how AI should work and figure out which ones actually survive contact with a production system, which in my current role tends to mean a system that is also being actively probed by someone who would rather it didn’t work.
These days my time splits between two modes. In one, I’m shipping. I take large language models, retrieval systems, and agentic scaffolds into production, and find out exactly where the abstractions leak (and, in a security setting, where an attacker might pry them open on purpose). In the other, I’m reading, writing, and thinking about modularity and compositionality as organizing principles for the next generation of AI systems. This site is where the second mode lives, though the two keep bleeding into each other, which is usually where the interesting questions come from.
Cybersecurity has turned out to be a surprisingly productive lens on the modularity questions I already cared about. Adversarial settings are unforgiving about sloppy abstractions, and the places where a learned system’s modules aren’t really modular tend to be exactly the places an attacker can reach through.
Interests
- Modularity and compositionality in learned systems
- Mechanistic interpretability, and what “structure” in a network even means
- Retrieval-augmented generation and its failure modes
- Agentic systems, and the design of interfaces between modules
- AI for cybersecurity: detection, triage, and autonomous response
- Cybersecurity for AI: prompt injection, adversarial robustness, and the failure modes of composed AI systems under active attack
- Causality, novelty detection, and out-of-distribution generalization
- Distributed systems, which I’ve come to think resemble neural networks more than is usually admitted
Elsewhere
- GitHub for code and half-finished experiments
- Twitter for shorter-form thinking
- LinkedIn for professional history
If you want to reach me about anything on this site, the easiest way is a direct message on Twitter or an issue on the relevant GitHub repository. I am especially interested in disagreement. Pointers to work that contradicts something I’ve written, or a good argument for why a research thread I’m pursuing is the wrong one, are genuinely welcome.