Open Haus Berlin: An evening for developers building in the open

An evening for developers building in the open.What does it actually take to run open models in production for millions of users?Engineers from n8n, NVIDIA, and Together AI share what they’ve learned building and operating real systems with open models. We’ll get into the practical engineering decisions behind model selection, evals, inference and serving, latency and throughput, reliability, and scaling workloads as usage grows.Join us in a digital art museum for an evening of meeting like-minded builders, learning from engineers making open models work at scale in real production systems, good food and even better vibes.Agenda5:30 p.m. – 6:00 p.m. Arrival, drinks & networking6:00 p.m. – 6:45 p.m. Three talks on building with open frontier models Scaling context in production LLM systems: Riqwan Thahamir, Staff Engineer, n8n As LLM applications grow more complex, managing context efficiently becomes critical to performance and cost. In this session, Riqwan Thamir, Staff AI Engineer at N8N, shares practical strategies from N8N's own infrastructure, including context engineering for multipurpose yet efficient agents, managing prefix cache, dynamic tool loading, and progressive context and skill loading. Attendees will walk away with concrete techniques for building systems that scale without sacrificing speed or reliability. Agent-Aware KV Caching: Harry Kim, Principal Product Manager, NVIDIA AI agents spawn sub-agents, compact their context and run long multi-turn sessions, so their KV cache patterns are hard to predict. This talk shows how the Dynamo router and the KVCR library work together to share KV cache across workers, recover it when an engine fails instead of recomputing it, and use signals from the agent harness to keep the KV that matters and evict the rest. LLM inference for the era of agents, Zain Hasan, Staff AI/ML Engineer DX, Together AI Every time an agent takes a step, it sends the entire conversation so far back to the model as a brand-new request to the inference engine. In this talk we'll start with how a single request gets served, from generating tokens one pass at a time to the KV cache, prefill and decode, continuous batching, pre-fill decode disaggregation, and PagedAttention. Then we'll look at what changes with agents: long prompts with short outputs, hundreds of turns per task, and sessions that go quiet while a tool runs, and how techniques like prefix caching and cache-aware routing keep all of that fast. 6:45 p.m. – 8:15 p.m. Bites + conversations with fellow builders8:15 p.m. Last call8:30 p.m. See you next time