AI Compute Architect

AI Compute Architect
📍 Medellín, Colombia

About the Role
We're building a next-generation AI infrastructure platform that delivers high-performance bare-metal GPU compute for large-scale AI training and inference.

As our AI Compute Architect, you'll lead the architecture, qualification, and operational readiness of our GPU infrastructure across rapidly scaling data centers. You'll make the platform work as an integrated production system that connects GPU servers, firmware, OS, NVIDIA software, high-performance fabrics, storage, provisioning, scheduling, observability, and on-site operations.

You won't replace our networking, storage, facilities, DevOps, or site-ops specialists, and you won't build every automation pipeline yourself. Your job is to provide the technical model that connects these domains. You'll define requirements, interpret the evidence, find where failures originate, and decide when infrastructure is ready for production. This is a hands-on architecture role, close to the hardware but with a platform-wide view.

What You'll Do

• Define the architecture: Set lifecycle standards for NVIDIA HGX platforms and validated baselines for firmware, BIOS, BMC, OS, drivers, CUDA, and NCCL. Own compatibility requirements and the model of GPU/CPU/memory/PCIe/NVLink/NVSwitch/NIC/storage relationships. Decide what testing each platform change requires before production.

• Lead qualification and burn-in: Design the strategy for new clusters, new GPU platforms, repaired nodes, and major changes. That includes load profiles, duration, coverage, and pass criteria. Analyze fleet-wide distributions and outliers to separate isolated failures from systemic issues in hardware, firmware, software, topology, power, thermals, or test configuration.

• Diagnose health and performance: Interpret NVIDIA diagnostics and telemetry, covering GPU, NVLink, NVSwitch, PCIe, memory, network, thermal, and power signals. Lead root-cause analysis for intermittent, multi-node, and fleet-wide issues, including collective and point-to-point communication problems.

• Validate with real workloads: Define representative training and inference workloads, including dense, MoE, and different parallelism strategies. Set reproducible baselines for throughput, scaling, latency, and numerical behavior, and pinpoint whether degradation comes from compute, communication, storage, scheduling, or configuration.

• Integrate across domains: Work with Network (topology, GPU-to-NIC locality, RDMA), Storage (access, checkpointing, recovery), Platform/DevOps (scheduler, containers, provisioning), and Facilities (power, thermal correlation). Route issues to the right specialists with solid evidence.