Jobs at Sciforium

The GPU Kernel Engineer will design and optimize GPU kernels for large-scale AI systems, integrating them into ML frameworks and collaborating across teams to enhance performance.

Develop and pre-train large byte-native and multimodal foundation models. Implement architectures, training objectives, optimization methods, and stable scaling recipes. Build production-grade training infrastructure, run ablations, analyze training dynamics, and improve model quality, efficiency, reliability, and scalability across GPU-based distributed environments.

Develop the core systems for multimodal AI models, focusing on building model serving platforms and enhancing developer experience with tools and APIs.

Develop the model serving platform for multimodal AI models, optimizing performance, collaborating with researchers, and ensuring reliable production infrastructure.

Lead advanced AI/ML research initiatives, developing novel algorithms and optimizing large-scale training systems. Mentor junior researchers and produce impactful research outputs.

The role involves optimizing and maintaining the software stack for large-scale AI training, focusing on CUDA/ROCm, JAX, and PyTorch to ensure efficiency and stability in distributed training systems.

Design, deploy, and operate the complete networking stack for large GPU clusters: RDMA fabrics (InfiniBand/RoCE), frontend/storage networks, security, cross-site/cloud connectivity, automation, telemetry, and performance debugging to ensure the fabric never bottlenecks training and inference workloads.

Own physical health and infrastructure of GPU clusters: respond to hardware outages, monitor GPU thermals/power, coordinate vendor RMAs and data center work, rack and bring up nodes, maintain Linux OS across bare-metal servers, secure networking and SSH, and manage directory and network storage services to keep compute environment stable for research and product teams.

Owner of the GPU cluster software stack: build/version OS images, drivers, and provisioning pipelines; automate validation, upgrades, and self-healing; operate GPU-enabled Kubernetes and Slurm; maintain driver/runtime compatibility and tuned PyTorch/JAX environments; debug deep accelerator and networking issues and run observability/benchmarking tooling.