About Me

I am currently a postdoc with the inimitable Vyas Sekar at Conviva and a visiting researcher at CMU. I graduated with a PhD from UIUC, where I was fortunate to be advised by Brighten Godfrey. My thesis looked at fundamental techniques to diagnose failures in networked systems.

I am currently working on building accurate and explainable data analysis agents, failure diagnosis in networked systems and network topology. I like to think about problems at the intersection of systems and theory.

My co-pilot.

Projects

Failure diagnosis in networked systems

• MoCE: A Mixture of Context-aware Experts Framework for Troubleshooting Internet-scale Services [NSDI 2026]
MoCE leverages an ensemble of algorithms to detect various failures modes. It provides a programmatic framework for developers to succinctly express their RCA logic using MoCE's DSL.


• Automatically Surfacing Opportunities for Improvements In Internet-Scale Applications [HotNets 2025]


• Murphy: Performance Diagnosis of Distributed Cloud Applications [SIGCOMM 2023]
How do we diagnose the reasons for performance issues in cloud apps? Murphy incorporates entity relationships via a Markov Random Field (MRF) for accurate diagnosis.
• Also see VMware's blog for product adoption.


• Flock: Accurate Network Fault Localization at Scale [CoNext 2023]
Probabilistic Graphical Models (PGMs) are the most accurate way to localize network failures, but are very slow. Flock employs new algorithms to accelerate inference via discrete PGMs at datacenter scale.

Datacenter network topology

• Starfish: A Topology-Routing Co-Design for Small-Scale Data Centers [NSDI 2026]
We propose new topologies for small scale datacenters that match the performance of expanders while being as intuitive as Clos topologies. We designed the first practical routing scheme for flat topologies, based on VRFs.
• Related: New AWS datacenters will be based on flat topologies with VRF-based routing
• Earlier workshop version: Spineless Data centers [HotNets 2020]
  

Distributed tracing

• TraceWeaver: Distributed Request Tracing for Microservices Without Application Modification [SIGCOMM 2024]
TraceWeaver recreates request-traces using a reconstruction algorithm based on timing information, eliminating the need to instrument applications for passing request headers.

Large-scale parallel sorting algorithms

• Optimal Round and Sample-Size Complexity for Partitioning in Parallel Sorting [SPAA 2023]
We prove tight lower and upper bounds for determining balanced splitters for parallel partitioning and sorting.


• Histogram Sort with Sampling (HSS) [SPAA 2019]
HSS is a large-scale parallel sorting algorithm that combines sampling and histogramming. HSS is faster than alternatives, both practically and theoretically.