Internships on a 96 GB card.
Our rack, upstate: one RTX PRO 6000 with 96 GB and a DGX Spark, in a twenty-unit open frame. Live projects. Each track ends with a report or build guide with your name on it.
Four tracks.
GPU systems
The Blackwell box is yours for a scheduled block. Run the benchmarks that never fit into project time. Profile kernels with Nsight. Thermal behaviour under sustained load is part of the work.
You leave with a published technical report.
ML research
Work inside the fine-tuning pipeline. Run ablations on LoRA rank and alpha, or score a model family against the NTQS harness. There is usually a real corpus in the queue, and client data stays put.
You leave with co-authorship on the write-up.
Infrastructure
Build a reference server from the parts list up. You assemble it, burn it in under sustained load until it holds, then write the guide. Some of these designs reach customers.
You leave with a hardware build guide with your name on it.
Deployment & solutions
Carry a finished model onto a customer’s machine and leave it running: containers, auth, logs wired into their stack. Then a handover with the people who will operate it. You write the runbook.
You leave with a deployment runbook of your own.
The bar is reasonable.
- You want to know how it works. Having touched a GPU before helps. Wanting to understand one matters more.
- You finish the thing. Every track ends in a deliverable somebody outside the building can read.
- You say when you are stuck. Two days in circles is useful information. Say it early and it is usually a ten-minute fix.
Apply for a track.
Tell us which track, one thing you have built or taken apart, and when you are free. A paragraph is enough. We reply within two working days.