Hannah B. Pasandi
Research Scholar Fellow | UC Berkeley Sky Computing Lab
Research Scholar Fellow
Research Scholar Fellow | UC Berkeley Sky Computing Lab
I am a wireless and networked systems researcher currently serving as a Research Scholar Fellow at UC Berkeley's Sky Computing Lab, where I contribute to the SkyPilot project with Professor Ion Stoica. I have been fortunate to work with exceptional researchers throughout my journey: Prof. Tamer Nadeem at Virginia Commonwealth University during my PhD, Prof. Sylvia Ratnasamy and Prof. Scott Shenker at UC Berkeley's NetSys Lab, Dr. Juan A. Fraire and Prof. Hervé Rivano at INRIA Lyon, Prof. Franck Rousseau at LIG Lab (Université Grenoble Alpes), Prof. Gil Zussman at Columbia University's Wireless and Mobile Networking Lab, and industry leaders at Qualcomm, Skylark Wireless LLC, and Bell Labs.
Our everyday life increasingly runs on two intertwined infrastructures. Networks carry our calls, our work, and our payments; AI agents are beginning to act across them on our behalf. I design and build systems that keep both trustworthy, predictable, accountable, and safe by design. Such an objective is fascinatingly challenging and complex, since real deployments are chains of layers, radio schedulers, satellite links, serving stacks, agent plans, that look independent yet carry hidden dependencies that must be understood, captured, and acted on; otherwise a guarantee that holds at one layer quietly collapses at the next, and in these systems a quiet collapse is a safety event, a dropped emergency call, a drone flying on a stale plan, an agent committing an unverified action. My research draws from networking, distributed systems, machine learning, security, and formal methods, model checking, SMT-based verification, and runtime certificates, a toolkit I first trained in during my master's in Grenoble.
My research interests are network protocol synthesis and certification, LLM serving and scheduling across device, edge, cloud, and direct-to-cell satellite tiers, plan-level accounting and reliability for agentic AI, cellular and voice privacy, and the security of connected vehicles and UAVs.
The commitment underneath the work is that systems should make decisions on predicted future state and prove bounds against the prediction error, rather than react to measured present state and bound average-case performance.
One principle, two arenas. In networks, teNET deploys only protocol compositions that carry a certificate they cannot collapse, and OrbitMesh turns direct-to-cell satellite into a first-class serving tier for LLM requests. In AI agents, where the unit of work is now the plan rather than the single call, Patient, Confine, and Gauge schedule, route, and evaluate at the plan level. Loom is both arenas in one system: an LLM agent held inside a safety certificate while it composes a live wireless MAC layer.
The same discipline extends to connected vehicles and smart farming through the students I advise; meet them, and everyone I am lucky to work with, on the People page.
Some of my work has shipped into production. At Skylark Wireless, my MU-MIMO scheduler improved spectral efficiency by 23 percent and serves millions of users on commercial 5G fixed-wireless deployments. At Qualcomm, my Wi-Fi 7 protocol contributions are integrated into commercial chipsets. My open-source tools (SIMSAT, SATMetrics, OffloadSim) are used by satellite operators and academic groups evaluating LEO constellations.
Wireless protocols across Wi-Fi 4 through Wi-Fi 7 and Wi-Fi 8 drafts. Cellular and 5G NR across 3GPP Releases 15 through 18, with O-RAN Near-RT RIC rApp development. Direct-to-cell LEO satellite systems across Starlink, Kuiper, OneWeb, and Globalstar architectures. Multi-tier LLM serving across device, edge, cloud, and satellite compute. SkyPilot, vLLM, Triton, and Ray for inference orchestration. PyTorch and JAX for learning. StarryNet, ns-3, and MATLAB for protocol simulation. C, C++, Python, Rust, and Go for systems code.
♠ A note on my name: I go by Hana or Hannah. My publications appear under Hannaneh B. Pasandi and Hannaneh Barahouei Pasandi.
♠ Collaborations: If you want to build systems for networks or AI agents, feel free to reach me at h.pasandi@berkeley.edu. My inbox has a soft spot for hard problems.
My work has been recognized by the N2Women Top 10 Rising Star Award, the NCWIT Collegiate Award, and the Heidelberg Laureate Forum, and supported by the Schlumberger Foundation and HIDA (Germany).
Nominated to join Sigma Xi, the research honor society whose members since 1886 have included Albert Einstein, Richard Feynman, and more than 200 Nobel laureates.
Our paper "The Reasoning Tax: Profiling Test-Time Compute Carbon in Agentic AI Workloads" is accepted to HotCarbon 2026.
Our VIN2VICTIM paper, in collaboration with Rivian, was accepted at USENIX VehicleSec 2026.
Our paper "CacheCatalyst" has been accepted to USENIX NSDI 2025.
Our paper "Balancify" won Best Paper Award and Best Contribution Award at ACM Student Workshop CoNEXT 2025.
Co-chaired ACM CoNEXT Student Workshop and gave a talk on "AI for Networks and Networks for AI".
Our three papers were presented at MobiCom'25.
Invited as Young Researcher to the 10th HLF in Germany.
Research on datacenter microbursts accepted to ACM SIGMETRICS 2025.
Received Best Paper Award for "Role of Machine Learning in Satellite IoT".
Talk on "Privacy and Security Challenges in Space Terrestrial Networks".
Delivered talks at IEEE S&P, USENIX SOUPS/Security, Microsoft BlueHat, and MSR on 5G privacy.
Received grant from Helmholtz Information and Data Science Academy, Germany.
Eight lines of work. Tap a thrust to open it.
Serving recomputes overlapping context across requests instead of reusing it; the cache that could carry the load is thrown away.
Per-call metrics grade agentic systems healthy while whole plans fail. Nothing in the stack accounts at the level of the plan the user asked for.
An agent that dies mid-plan cannot tell which external effects already landed, so restart means redo or double-do.
Learned systems assume datacenter-grade GPUs. The commodity fleet the world actually owns fails in ways training loops do not survive, priced with production failure data from Meta's Llama 3 Herd report.
Agent plans are scheduled one call at a time. A plan-level budget covers the retries, verification, and waste that single-call scheduling never sees. Full version in progress.
Failure labels for agents assume annotators agree on what broke. Measured agreement varies by failure mode, and routing needs reliability contracts that survive the disagreement.
Agent steps stall waiting for state the plan already announced. PALSE reads the plan while generation is in flight and stages what the next step needs at the edge.
Inference serving stops at the cloud edge while direct-to-cell satellites sit overhead as compute nobody schedules. A four-tier serving path for agent workloads includes the satellite.
A room of idle Apple silicon sits beside a phone paying cloud latency for every agent token. Tether routes inference across the personal device mesh.
Reasoning tokens never appear in carbon accounting; test-time thinking is off the books.
Safe RL still learns by violating. Systems that cannot afford a single violation cannot use it. VETO acts only when a verified ensemble agrees the action is safe.
No theory predicts when cooperating learned agents snap under distribution shift; operators discover the phase transition by outage.
Multi-tier LLM serving reacts to 5G handovers and thermal throttling after they hit. At mobile timescales reaction is too late; prediction has a thirty-second window.
RL for rate and power control learns by violating the limits the RAN exists to hold. A decision-time mask removes violating actions from play, so it learns clean.
Direct-to-cell LEO handovers strike mid-session with no warning, at orbital cadence. An O-RAN Near-RT RIC rApp predicts them with a GBM and LLM pair.
Satellite IoT uplinks transmit blind into poor passes, spending the scarcest link budget there is. Pass-aware deferral waits for the right window across two orbits.
A LEO constellation crosses jurisdictions every few minutes. No mechanism keeps data flows legal while they move.
Mobile LLM requests route on latency alone while carbon differs by an order of magnitude across device, edge, and cloud. Full version in progress.
Cellular signaling includes a path that can quiet a device. The paper asks who can trigger it and at what cost.
A network can only act on what it sees early enough. Lead Time measures how far ahead predictions land against the cost of acting on them.
One trained predictor turns sparse mobility signals into persistent tracking.
Vehicles navigate on a signal anyone can forge. With Rivian, detecting GPS spoofing before the planner believes the lie.
MAC layers freeze at standardization time and conditions do not. Loom lets an LLM agent recompose a live wireless MAC while a safety certificate keeps every action inside proven bounds.
Learned MAC controllers act unmasked, so one bad action reaches the air. A safety mask filters actions before transmission.
Protocol synthesis emits code, not guarantees. Arachne composes MAC protocols from verified blocks, and every composition carries an interference-safety proof.
Agents propose changes to systems they should never control directly. Vouch splits proposer from deployment license; a small certificate, not the model, decides what may run.
Networks treat LLM traffic as one workload. Measured at packet level, it is three, and designs tuned for chat mis-serve the agentic traffic growing fastest.
RAN automation is adopting LLM agents with no effect contract and no plan-level accounting. The RIC is about to run agents blind.
View all publications on Google Scholar →
Interested in collaboration, research discussions, or speaking opportunities?