Training that survives the real world
Decentralized and fault-tolerant training, performance-modeled scheduling across clouds, clusters and spot markets. Rust/Go/Python, real GPUs, real failures.
Careers
We're a small founding team in Frankfurt building every layer of applied intelligence — the GPU fabric, the data engines, the models, and the runtimes that put them on chips. Small team, whole-stack ownership, hardware on your desk.
What you'd work on
Decentralized and fault-tolerant training, performance-modeled scheduling across clouds, clusters and spot markets. Rust/Go/Python, real GPUs, real failures.
Sector-specific LMs and VLMs: training recipes, distillation and quantization that keep accuracy when the parameter budget drops two orders of magnitude.
PanoComp and PanoRT — kernel generation, memory planning and a lean runtime for small-scale accelerators that datacenter toolchains ignore.
Multi-tenant platform, dashboards, CLI/SDK — the surface through which every customer touches everything above.
How we hire
We don't post laundry-list job descriptions — we hire when we meet people who can't stop pulling on one of the threads above. Send us what you've built (repos, papers, write-ups beat CVs) and which layer of the stack you want to own.