AI for Research NemoClaw Lab Workshop

​Date:

​2nd October 2026 (Friday), 90 minute session.

​The same workshop will be held four times at 9:00, 11:00, 13:30 and 15:30 to accommodate more attendees.

​Location:

​Inox, Students' Union Building, The University of Sheffield, Western Bank, Sheffield, S10 2TG

​90 minutes, hands-on, instructor-led. You will be working in a team of two at a shared Dell Pro Max with GB10.

​AI has changed shape. A chatbot is one prompt, one answer and a couple of thousand tokens. An agent is a loop: plan, call a tool, read the result, decide what to do next, repeat. The same task now runs to 50,000 or 500,000 tokens, takes minutes rather than seconds, and costs something different every time. On per-token cloud pricing that turns pennies into dollars per task, makes forecasting guesswork, and leaves you asking where your data goes, what the agent is allowed to touch, and what happens when you hit rate limits at peak hours.

​In this lab you find that out by doing it. Your team gets a Dell Pro Max with GB10 system running DGX OS, set up and tested before you arrive. You stand up the NVIDIA NemoClaw runtime and see what each piece of the stack does: vLLM serving the model locally on the GPU, OpenClaw running the agent loop, NVIDIA OpenShell sandboxing the whole thing. You then wire an agent up with tools, memory and a real workflow, and run it, watching tokens, cost and latency in real time while you see how governance, orchestration and placement affect what the agent can do. Then you connect that local agent to data-centre AI services, so you leave knowing how the local and the data-centre pieces fit together.

​

​Have a look before you sign up:
Overview · Developer docs · GitHub

​

​Prerequisites

​You should be comfortable in a Linux terminal (we are on DGX OS, Ubuntu 24.04), know roughly what Docker does, have used an LLM through an API or a coding agent, and be able to read a short Python script and change it. You will not be writing an application from scratch. You also need to be there in person for the full 90 minutes and be happy to share a GB10 with another participant.

​Nice to have, not required: vLLM or another local inference server, prior OpenClaw use, hands-on time with a GB10.

​Everything else is provided. Hardware, OS, model and lab environment are pre-installed. Bring nothing but yourself.