I study isolation in cloud systems that many customers share: when one tenant can read or affect another's data, how to test for that, and what correct isolation costs. I've worked in backend and platform engineering since 2021.
Papers (single-author):
- Accepted: a verification protocol for silent tenant-isolation failures (Advances in Cyber-Physical Systems).
- Under review: runtime tenant onboarding and per-tenant connection pools as TTL caches.
- In preparation: stale adapter state in LLM prefix caches. In vLLM 0.31.0 with runtime adapter loading enabled, after a LoRA adapter update through the stock API, 771 of 780 probe requests were served prefix tokens computed before the update (related crash, fixed upstream).
Retry storms: I reimplemented the queue model from "Modeling Metastability" (HotNets '25) and derived a condition for when it can get stuck after a load spike has passed.
Merged upstream: 23 pull requests in 16 projects, for example:



