Blog

ForkOps at SWGeno 2026 and Whose Oracle Is It? at APR 2026, both workshops co-located with ASE in Munich. Generated software needs fleet operations, and generated patches need an authority problem solved.

Two papers at AgenticOS 2026 and one at PACMI 2026, both workshops co-located with SOSP in Prague. And I will be giving the keynote at AgenticOS. If you are attending either workshop, come say hi.

Two new papers with Quentin Botha and Christoph Siebenbrunner at BCCA 2026 in Barcelona. Agent Flight Recorder gives long-horizon agents a tamper-evident audit trail anyone can verify, and Bonded Recourse settles the harm an agent leaves behind through smart contracts.

Everyone is predicting that AI will flatten the software stack into one English-speaking agent, until nothing of the old software remains. It will not go that far. Complexity is conserved, the end-to-end principle still collects its debt, and the stack does not flatten all the way down. It sorts itself by what you can afford to trust.

New paper with Akanksha Narula and Mofasshara Rafique at ICML 2026 in Seoul. Incidental prompt cues silently decide which algorithm an LLM ships, even when every output passes the same tests.

I am giving the keynote at EuroMLSys 2026 in Edinburgh. A short teaser on what happens to software when code becomes disposable.

New paper with Mofasshara Rafique at EuroMLSys 2026 in Edinburgh. Classical virtual memory meets agent state management.

New paper at ARCS 2026 in Mainz. We decouple agent planning from infrastructure actuation so that LLM-based remediation agents cannot cause more damage than the incident they are trying to fix.

Akhil Arora and I are running a half-day tutorial at SOSP 2026 on building and hardening runtimes for long-lived, tool-using agents.

When intent became the ultimate abstraction and code stopped being art. A future-looking backward scenario set in 2029.

When code becomes cheap to generate but expensive to trust, architecture shifts to minimize what must be trusted and maximize what can be verified. Three concepts explain why: the Rehydration Law, the Forkability Trilemma, and the Intent-Compilation Theorem.

My paper on using LLMs to canonicalize SQL queries for semantic caching in OLAP systems has been accepted at DOLAP 2026.

I released scode, an open-source tool that sandboxes AI coding harnesses — blocking access to your credentials, personal files, and sensitive directories. One bash script, works with any tool.

Saumya and I talked about LoRACode on a Ploutos AI live stream — how LoRA adapters make code embeddings better without retraining the whole model.

OpenClaw has taken the AI world by storm with 150,000 GitHub stars. What makes it work? Two simple abstractions that Apple, Google, and Microsoft somehow missed. A systems perspective on Peter Steinberger's elegant design.

A year ago I submitted an ambitious Emmy Noether proposal on modular AI systems. It was rejected. But watching the field evolve since then, I cannot help but feel that we were right all along.

As we close out 2025, it is clear that building reliable AI systems is fundamentally a systems problem. Context engineering, state management, and failure handling matter as much as the models themselves.

Our multi-agentic Text-to-SQL framework will be presented at the Deep Learning for Code workshop at NeurIPS 2025.

Our paper on efficient and reproducible LLM inference has been accepted at EMNLP 2025 in Suzhou, China.

Our research on requirements engineering for decentralized systems will be presented at RE4Web3 2025 in Valencia, Spain.

I have been honored with an Associate Fellow appointment at the University of Saarland.

Our comprehensive survey on graph neural network partitioning strategies will be presented at GRADES-NDA 2025 in Berlin.

Our work on parameter-efficient fine-tuning for code embeddings will be presented at the Deep Learning for Code workshop at ICLR 2025.

Our work on enabling efficient and secure FPGA acceleration for serverless functions will be presented at HPDC 2025.

Our framework for endowing LLMs with stateful capabilities will be presented at EuroMLSys 2025 in Rotterdam.

Our work on making LLM-enhanced database queries practical through query-specific optimization will be presented at DOLAP 2025.

Our work on reliable latent knowledge estimation in LLMs will be presented at WSDM 2025 in Hannover, Germany.

Co-teaching the undergraduate Operating Systems course at UdS in Spring 2024.

Our dedicated blockchain node for analytics will be presented at EdgeSys 2024 in Athens, Greece.

Launching a new seminar on the systems challenges of deploying and scaling LLMs.

Our demo on generating realistic database benchmarks from query logs will be presented at ICDE 2023.

Co-teaching the undergraduate Distributed Systems course at UdS in Spring 2023.

Starting a new chapter leading the Data Systems Group at the Max Planck Institute for Software Systems.

Our work on extracting vehicle trajectories from low-quality traffic cameras has been published at CompAuto 2022.

Addressing challenges in evaluating machine learning-enhanced data management systems and proposing new benchmarking approaches.

Introducing Tesseract, the first distributed system for executing general graph mining algorithms on evolving graphs.

Our workshop paper on benchmarking learned database systems has been accepted at SMDB 2021.

Tesseract, the first distributed system for graph mining on evolving graphs, has been accepted at EuroSys 2021.

EPFL has published my doctoral dissertation on load balancing in cluster computing applications.

Starting a postdoctoral fellowship at MIT's Computer Science and Artificial Intelligence Laboratory.

The Hailstorm codebase is now publicly available under the Apache 2 license.

Our Hailstorm presentation received the best video award at ASPLOS 2020.

Presenting Hailstorm at ASPLOS 2020 — a system for improving load distribution in distributed LSM-based databases.

Hailstorm has been accepted to ASPLOS 2020 in Lausanne, Switzerland.

The Hurricane source code is now publicly available under the Apache 2 license.

Chaos achieved #1 ranking in Graph500's capacity benchmark for single-source shortest path, processing 128 billion edges on just 8 machines.

Presenting Hurricane, a high-performance distributed analytics system designed to handle data skew through adaptive work partitioning.

Our paper on Hurricane has been accepted at EuroSys 2018 in Porto, Portugal.

Our team's short film 'Jenna and Danbot' exploring AI relationships won the public prize at the Exposure Science Film Hackathon.

Chaos achieved 4th place in Graph500's capacity ranking, processing 16 trillion edges on just 20 commodity servers.

Presenting Chaos, a graph processing system for analytics on big graphs using small clusters.

The Chaos graph processing system is now open source under the Apache 2 license.

The X-Stream graph analytics system is now open source under the Apache 2 license.

Our paper on Chaos has been accepted at SOSP 2015 in Monterey, California.

Presenting X-Stream, a system for analyzing large-scale graphs on a single machine through storage-centric design.