Peter Pietzuch

Agent backtracking and parallel tree-search become practical across real-world tools when incremental local checkpoints are paired with automated remote compensating actions, driving a 15x task quality gain at just 3% overhead.

Achieving up to 2.3x higher training throughput, Tangram transforms how we approach parallelization in heterogeneous GPU clusters.

Scepsy achieves up to 2.4x higher throughput and 27x lower latency for agentic workflows by intelligently scheduling LLMs based on aggregate execution time shares, outperforming systems that treat LLMs independently or rely on manual allocation.

Nanvix achieves 20-100x higher serverless deployment density by decoupling ephemeral and persistent state, outperforming state-of-the-art systems.