- Lead netFound, a multi-modal hierarchical foundation model pretrained on 1 TB of raw network traces.
- Built NetBurst, an event-centric time-series foundation model that beats Chronos in traffic forecasting by 2.07×–200×.
- Co-authored work on interpreting network foundation models (NeurIPS 2025) and the Pramana data-generation backend (HotNets 2026).
Satyandra Guthula
I'm a Ph.D. student in the Systems and Networking Lab at UC Santa Barbara, advised by Prof. Arpit Gupta. I work on representation learning for network data. My current focus is NetBurst, an event-centric foundation model for bursty, intermittent time series such as network traffic. Earlier, I built netFound, a foundation model pretrained on terabytes of packet traces for network security tasks. In summer 2026 I interned at Google, where I applied time-series foundation models to diagnosing faults in optical networks.
Before my Ph.D., I spent four years at VMware building the data-collection and analytics pipeline for vRealize Network Insight, a data center network monitoring product. That experience shapes how I work: I care about models that hold up on production data and survive deployment.
News
- Pramana, a composable backend for empirical networking research, will appear at HotNets 2026.
- Continuing at Google as a Student Researcher (remote, Sunnyvale), adding physical fault localization to the diagnosis pipeline.
- Software Engineering Intern at Google, working on time-series diagnosis for optical networks.
- NetBurst, an event-centric foundation model for bursty time series, is on arXiv.
- Demystifying Network Foundation Models accepted at NeurIPS 2025.
- Returned to ESnet as a research intern, working on time-series foundation models for IPFIX traffic.
- Research intern at ESnet (Lawrence Berkeley National Lab), deploying netFound on production data.
Research
Network data is abundant but labels are scarce, and most ML-for-networking models are trained for a single task on a single dataset. My work asks whether we can pretrain once on large volumes of unlabeled traffic and reuse the result across many tasks and networks. Current threads:
- Foundation models for packet traces. netFound learns hierarchical representations of packets, bursts and flows from 1 TB of raw traces.
- Bursty, intermittent time series. NetBurst models traffic as a sequence of events, which suits the sparse, heavy-tailed signals that general-purpose forecasters handle poorly.
- Evaluating what these models learn. Probing pretrained network foundation models without relying on downstream tasks.
Publications & Preprints
Full list on Google Scholar.
Patents (VMware)
Experience
Software Engineering Intern · Bengaluru, Summer 2026
- Now extending the diagnosis pipeline with physical fault localization, using network topology for spatial context.
- Built a retrieval-based pipeline for diagnosing optical-network faults from time series, using time-series foundation model embeddings and clustering.
- Sent only low-confidence cases to an LLM, which made continuous monitoring practical at much lower latency and cost.
- Built a time-series foundation model for sparse, heavy-tailed IPFIX traffic: 10× lower MSE than Chronos and 94% application-classification accuracy.
- Deployed netFound on ESnet production data; distillation halved parameters at under 5% accuracy loss. Scaled training on NERSC with PyTorch DDP and Spark.
- Owned the network domain of vRNI: data collection from switches, routers and firewalls, and VM-to-VM path stitching. Mentored four new graduates.
- Redesigned the telemetry pollers, cutting memory from 12 GB to 1 GB and raising throughput 8×.
- Built an in-memory emulator of the product pipeline that became a CI gate on every pull request.
Education
Contact
satyandra@ucsb.edu
Department of Computer Science, Harold Frank Hall 5112
University of California, Santa Barbara, CA 93106