I am a Senior Researcher and Lecturer at ETH’s Systems Group working with Ana Klimovic. I am co-teaching a seminar on Systems for AI . My intro slides summarizing the fast-moving space are available here.
My research is centered around efficient AI and data systems, focusing on how large-scale AI workloads, such as retrieval-augmented generation and agent-based inference, can be served reliably and efficiently under real-world constraints. My work emphasizes principled abstractions, declarative system design, and the identification of fundamental trade-offs that shape scalable, trustworthy AI infrastructure.
Previously, I was a Senior Researcher at Apple, a PostDoc at UZH (DaST with Dan Olteanu) and defended my thesis at ETH’s Systems Group with Ce Zhang.
If you are interested in collaborating, please reach out directly. I also welcome motivated students at ETH for projects and theses. To apply, send me an email with your CV and transcript of records attached.
Research Focus Areas
As large‑scale machine‑learning models become central to modern applications, the underlying systems must efficiently manage computation and intermediate state under tight latency and resource constraints. Our research focuses on the systems challenges of serving complex, sequential AI workloads, such as large language models and reinforcement‑learning pipelines, including secure sandboxed executions. We design distributed infrastructure and scheduling mechanisms that optimize the memory hierarchy and intermediate state (e.g., key‑value caching) while accounting for secure tool execution across heterogeneous hardware. By bridging the gap between model architectures and physical hardware constraints, we aim to improve the throughput, reliability, and cost‑efficiency of AI serving systems in production environments.
Modern AI systems increasingly operate dynamically, adapting their behavior based on continuous user interactions and tool utilization. Understanding how these systems perform in real‑world settings requires moving beyond static benchmarks and analyzing the execution traces generated during deployment. We investigate methods to model and predict the outcomes of these interactive workloads—such as anticipating resource requirements or determining optimal model‑switching strategies in real‑time serving engines. By developing systems that can learn from and act upon execution traces, we aim to improve the adaptability, reliability, and overall performance of interactive AI applications.
The reliability of AI systems heavily depends on our ability to accurately evaluate their performance, which is often obscured by noise from various sources in the deployment pipeline. Our research addresses the fundamental challenge of isolating true model behavior from artifacts introduced by infrastructure, evaluation sandboxes, biased metrics, and low‑quality data. We build tools and methodologies for automatic feasibility assessment, data‑quality analysis, and noise disentanglement to better understand the fundamental limits and actual system performance. By providing a more precise picture of system behavior under real‑world conditions, we enable practitioners to build more trustworthy and robust AI pipelines.
