Deep dives into the systems we build. No fluff, no marketing — just technical writing from the team in the trenches.
K8s was designed for stateless microservices, not 8-GPU nodes with NVLink topologies. Here's what we built instead — and why bare-metal scheduling outperforms container orchestration by 40% on inference throughput.
Read article →How we achieved 2.3x speedup on attention layers by fusing softmax, masking, and dropout into a single CUDA kernel. A walkthrough of our MLIR-based fusion pass.
Read article →When you have 10,000 AI agents coordinating across regions, strong consistency is a luxury you can't afford. We use operation-based CRDTs for state management — here's the architecture.
Read article →We trained a lightweight model to predict thermal throttling 90 seconds before it happens. By preemptively migrating workloads, we eliminated 99.7% of thermal-induced performance drops.
Read article →From network stack bypass (DPDK) to custom memory allocators, the full story of how we brought P99 latency from 12ms to 1.8ms for LLM token generation.
Read article →Why we started MetalBear, what problems we're solving, and our vision for the future of AI infrastructure. The manifesto for building things that never break.
Read article →Engineering updates, no more than once a month. No spam, unsubscribe anytime.