AI Infrastructure Engineer

About the Role

Spellbrush, the world’s leading generative AI studio behind niji・journey, is looking for an AI Infrastructure Engineer to join us in building out end-to-end ML infrastructure to run our models on all platforms.

What you’ll do

• Design, implement and run our next-generation inference architecture for running all our models powering all platforms and applications (mobile, web, etc.).
• Work alongside a fast-paced and nimble team developing the latest state-of-the-art image generation models serving over 16 million users.

Requirements

You have experience with large distributed systems

You have familiarity with the latest hotness like K8S, Kafka, NATS, Redis, etc. You’ve cut your teeth on both on-prem and multi-cloud clusters. But most importantly, you deeply understand the tradeoffs and the failure mode of each system you introduce (and potentially even have the battle scars to prove it!).

You have an excellent understanding of GPU’s handling large workloads

GPU workloads are different from traditional CPU workloads in very interesting ways. Experience deploying, or even optimizing them end-to-end, is a huge plus for this role.

The anime aesthetic resonates with you.

It's no secret – we're huge anime enthusiasts, and our work focuses on the anime aesthetic. Your work will enable millions of users to partake in an evolving creative movement.

You're comfortable working on small, fast-paced teams

We currently have a small tight-knit team on AI. You'll be working closely alongside some of the best AI researchers in the world, on the literal best image model in the world. We also believe in the unmatched speed of in-person teams, and prefer on-site collaboration in either our primary research office in Tokyo (downtown Akihabara), or San Francisco. Visa sponsorships are available.

Technology

At Spellbrush, we're pushing the boundaries of generative networks through our state-of-the-art diffusion text-to-image research. There's no shorted of interesting tech challenges here:

• Research: We make heavy use of Pytorch and JAX, and train large-scale models using both TPUs and GPUs. Our researchers spend each day developing and testing novel state-of-the-art diffusion models.
• Engineering: We make use of a lot of Elixir for real-time backend systems, Python for machine learning, and lots of simulation and game logic code written in C# via Unity. We also have tons of internal tools and apps written in Typescript, React, and RN.
• Infrastructure: We've done everything from building our own in-house GPU supercomputer, to scaling up massive inference clusters in our own datacenters, typically orchestrating via some unholy combination of SLURM, kubernetes, and NATS.