Sword Health Summit 2026
Conversational AI is moving into real clinical workflows. At the Sword Health Summit, Nebius and NVIDIA are running a hands-on workshop on how to test and improve AI agents before they reach real users.
Most teams evaluate agents on static benchmarks that miss how real conversations go wrong. In this session you’ll generate realistic synthetic conversations and use them to probe model behaviour, find failure modes, and improve your agent. You’ll use the NVIDIA Agent Toolkit to create the data and Nebius to deploy the models.
The workshop centres on NVIDIA NeMo UserSim, which simulates multi-turn conversations and produces performance reports for the LLMs under test. MindEval 2 serves as the evaluation probe, and a judge model fine-tuned by Sword Health scores the results.
Agenda
11:20–11:45 — From silicon to inference: Co-Designing the full stack for Scientific AI
Ilya Burkov, Global Head of Life Science and Healthcare Growth at Nebius
Join us to:
- See a live doctor–patient simulation. Two LLMs run a mental health consultation. The patient has hidden symptoms, and the doctor has to uncover them. You can nudge either side or take over a role yourself.
- Generate synthetic evaluation data with NeMo UserSim. Create realistic, varied conversations at scale and turn them into structured performance reports.
- Evaluate and iterate. Test model behaviour, refine prompts and tools, and measure whether your changes actually improve the agent.
- Work in a ready-to-go environment. Every attendee gets a preconfigured, one-click cloud workspace, with Nebius Inference serving the models under evaluation. There’s no setup and nothing to install.
14:00–15:45 — Build, simulate,evaluate: Testing conversational AI with NVIDIA agent toolkit and Nebius
Sr. Solution Architect at NVIDIA and René Schönfelder, Solutions Architect at Nebius
The next breakthroughs in healthcare and life sciences won’t come from bigger models alone. Scientific workloads push infrastructure in different ways. They involve massive multimodal datasets, heavy simulation and reasoning, fast-moving architectures, and a growing need to run models at scale, not just train them. Ilya makes the case for full-stack co-design, which brings silicon, systems architecture, models, and scientific workloads closer together, and shows what that makes possible: AI as a practical engine for discovery.
Centro de Congressos Alfândega do Porto
Try Nebius AI Cloud console today
