Selected workProduct ↔ systems
I selected these because each leaves something another person can inspect or reproduce: a test record, a configuration, a tool or a clearly stated limit.
Loading a large model is easy to mistake for proof that a platform is ready. This ongoing investigation checks the hardware baseline, runtime compatibility, model fit, correctness, performance, long-session behaviour and recovery separately. Each conclusion stays tied to the named machine, software and workload.
Read the ongoing seriesAn end-to-end system that imports Garmin Varia Vue front and Varia RCT715 rear footage over USB, builds continuous ride recordings, pairs usable views for review, and extracts a selected moment into a separate evidence share. Current limits are part of the project record rather than hidden.
Open the projectThis guide deliberately starts with disposable memory-backed storage. That proves the SPDK and RDMA path without putting a real drive at risk. The release is pinned, host access is explicit and the checks show where a lab configuration would need more work before production use.
Read the field noteThe risky assumptions are made visible before the configuration begins: which device is being exported, whether anything else is using it, which host is authorised and whether the entire RoCE path has been checked. The result is a persistent lab target with refusal checks around the most damaging mistakes.
Read the field noteI developed s3bench to test parallel byte-range downloads with controlled chunk sizes and concurrency. Discarding the data isolates the transfer path; writing it exposes the end-to-end path, including local storage. The tool does not assume that more concurrency is always better.