Nebius.Build/BER
The Nebius Build Tour is a series of in-person technical gatherings designed for engineers working on real-world AI infrastructure. Join ML engineers, platform architects, and technical founders for deep dives into deploying, scaling, and operating AI systems in production.
This is not a marketing conference — it’s a focused environment for builders to share architecture patterns, lessons learned, and practical approaches to running AI workloads at scale.
Nebius.Build/BER brings together engineers from Nebius and partners including Anyscale and Tavily, with sessions covering distributed training, running open source models in production, and scaling workloads across large GPU clusters. The program includes a mix of technical talks, partner sessions, and a hands-on workshop, where you can deploy your own model and connect it to a working inference pipeline.
The program includes talks on real architecture breakdowns and implementation details from teams running production systems, with discussion of inference scaling, latency, reliability, and cost-performance tradeoffs, alongside practical deployment workflows.
The content runs from 13:00 to 16:30, followed by a networking reception and a help bar, allowing time to connect directly with engineers and peers.
Register your interest to receive updates on upcoming events and be the first to secure a place when registrations open near you.
Why attend
Learn from engineers running production systems
Hear real architecture breakdowns and implementation details from teams deploying AI at scale
Go deep on infrastructure and performance
Explore topics like inference scaling, latency optimization, reliability, and cost-performance tradeoffs
Hands-on technical sessions
Participate in workshops and walkthroughs focused on real deployment scenarios
Participate in workshops and walkthroughs focused on real deployment scenarios
Who should attend
Full agenda
12:00-13:00 –– Registration and lunch
13:00-13:15 –– Opening remarks and scene setting
Waqas Makhdum, VP of Developer Relations & Community, Nebius
13:15-13:45 –– AI Cloud talk: From prototype to production: reliable clusters for large scale distributed training
Evgeny Arhipov, Head of Scheduler Services: Managed Soperator, Slurm, Nebius
13:45-14:30 –– Token Factory talk: Post training in production. Making open models reliable at scale
Dylan Bristot, Product Marketing, Token Factory
Sujee Maniyam, DevRel, Token Factory
14:30-15:00 –– Coffee break
15:00-15:25 –– Breakout sessions: select your preferred track on the registration form
Track 1 — Partner talk: Anyscale
Physical AI at scale: Efficient VLA training on hundreds of GPUs
Carl Winkler, Anyscale
Physical AI is entering a new era, with VLA models reshaping how robots learn to see, reason, and act from a single architecture. But scaling VLA training across hundreds of GPUs exposes a core infrastructure challenge: the data is heterogeneous, the compute is heterogeneous, and in traditional batch pipelines, expensive GPUs sit idle while CPU steps run.
This talk shows how streaming execution eliminates that waste by continuously feeding data through CPU and GPU stages in a single unified pipeline, keeping GPU utilization near 95% instead of waiting for each step to complete. We will walk through the full data pipeline, from show how fault-tolerant streaming on Anyscale, deployed on Nebius AI Cloud, lets teams efficiently utilize hundreds of GPUs, process massive multimodal datasets, recover from failures without restarting from scratch, and dramatically speed up the development cycle from data to training.
Track 2 — Workshop
Hands-on with OpenClaw: Bring your own model
Mikhail Rozkhov, TPM, Nebius
In this hands-on workshop, you will learn how to deploy a custom model and connect it to OpenClaw to build a usable AI service. We start by deploying a model to a serverless endpoint, making sure it is accessible, stable, and ready for inference.
From there, we integrate the deployed model with OpenClaw, configuring it to handle requests, route inputs, and expose a clean API. Along the way, we cover the practical details that matter in real setups: model compatibility, endpoint configuration, and how to make the system reliable end to end.
By the end of the session, you will have a working pipeline where your own model is deployed, connected to OpenClaw, and ready to be used in real applications.
15:25-15:50
Track 1 — Agentic Search: Context Management in Research Agents
Agentic search is not just about retrieving information. It is about delivering the right context, in the right form, at the right time for an agent to reason effectively. In this talk, we explore the architecture behind deep research systems through lessons from building Tavily’s research API. We examine how agent harnesses are designed, how context is managed across multi-step workflows, and which context engineering techniques most improve performance in practice.
Evan Rimer, Forward Deployed Engineer, Tavily
Track 2 — Workshop
Hands-On with OpenClaw: Bring Your Own Model (continued)
Mikhail Rozkhov, TPM, Nebius
15:50-16:05 –– Keynote address
16:05-16:20 –– Fireside chat with keynote speaker
Waqas Makhdum, VP of Developer Relations & Community, Nebius
16:20-16:30 –– Wrap up and closing remarks
Waqas Makhdum, VP of Developer Relations & Community, Nebius
16:30 –– Help bar
Networking happy hour with bites
Impact Hub Berlin
Try Nebius AI Cloud console today
