Engineering

The build log.

Benchmarks, architecture notes, and technical writing.

Good Practice and Silent Errors in Distributed Inference

Distributed inference fails quietly. Three practices -- joined observability, consistent container monitoring, and classified error states -- determine whether those failures are findable.

Read →

Open-weight model inference on dedicated UK hardware: benchmark results

Throughput and latency measurements across 23 instruct models, 8 embedding models, 3 text-eval models, and 7 TTS models. 738 completed jobs, zero errors, on private AWS infrastructure in eu-west-2.

Read →