Projects

Open each edge of a lattice independently with probability p. Above a critical value pc an infinite open cluster appears. But is there one exactly at pc? This is the θ(pc) = 0 problem of percolation theory. It was settled in two dimensions by Harris and Kesten, and in high dimensions by the lace expansion, but stayed open for dimensions 3 to 10 for decades. A September 2026 preprint posted by OpenAI claims a proof for every ℤd and every quasi-transitive graph with pc < 1. I made a 13-minute animated tour of how the argument works. ...

I coauthored an Anyscale technical deep dive on integrating NVIDIA Dynamo’s KV-aware selection machinery into Ray Serve LLM. The collaboration is a useful example of a systems component becoming more valuable when it can be embedded outside the system where it started. ...

I worked with Gcore’s AI team on the Dynamo side of a global inference-routing problem: when the same model is served in many data centers, how do you decide where each request should go without throwing away useful KV cache state? ...

I worked on DynoSim and coauthored the NVIDIA technical blog introducing it. DynoSim is a workload-driven discrete-event simulation of the NVIDIA Dynamo serving stack: a digital twin that combines engine timing, scheduler behavior, routing, planning, KV-cache effects, and workload traces on one virtual timeline. ...

Simulating spin glasses at scale has always been a pain — Python is too slow, and C++ is too painful. Peapods is my attempt at a middle ground: a Monte Carlo package for Ising spin systems where the hot loops are written in Rust and everything else stays in Python, glued together with PyO3. You just pip install peapods and go. ...

I worked on the Flash Indexer, a high-throughput global KV-cache indexer for NVIDIA Dynamo. The problem is to track which inference workers hold which KV blocks, and then answer routing queries fast enough that the indexer itself does not become the bottleneck. ...

Dynamo is NVIDIA’s open-source datacenter-scale distributed inference serving framework for generative AI and reasoning models. Built in Rust for performance and Python for extensibility, it supports disaggregated prefill and decode, dynamic GPU scheduling, and LLM-aware request routing across multi-node multi-GPU topologies. It supports backends including TensorRT-LLM, vLLM, and SGLang. ...

While working at BrainChip, I led the research and software-hardware codesign of TENNs (Temporal Neural Networks) — a family of deep state-space models designed for efficient, low-latency inference on edge hardware. This line of work developed in parallel with the broader SSM wave (Mamba, Gated Delta Net, etc.), cross-pollinating ideas around structured recurrences, parallel scans, and kernel parameterizations. ...

Unlike ferromagnetic spin models, a spin glass has a highly non-trivial phase transition profile. For mean field models (the SK spin glass), the glass experiences a replica-symmetry breaking phenomenon that is only fully understood recently. However, for short-range models, almost nothing is known rigorously, including basic problems such as the nature of the glass transition. A potential route towards settling such problems may rely on a suitable generalization of the random cluster model as a universal statistical tool. ...

The universality classes of equilibrium phase transitions of many statistical models are well known. On the other hand, very little is understood about the phenomenon of non-equilibrium criticality, including its existence. Nonetheless, from numerical simulations, it appears that memory effects are fundamental driving forces behind dynamic criticality. They are capable of establishing long-range order irrespective of the structure of the equilibrium model. A theoretical understanding of the role of memory in complex systems is important to understanding critical phenomenon as a whole. ...