ENGINEERING JOURNAL / COMPANION TO CVPART OF SHAYANRADI.COM
FIELD NOTES / VOL. 001
Ideas into
systems._
Writing about the distance between a model that works in a notebook and a system you can actually operate.
RASTER ID / FIG. 001 ◼
0x53 / 0x527 × 12
SHAYAN RADIFIELD INDEX / 001
ENGINEERING, NOT HYPE.REPRODUCIBILITY BY DEFAULT.LEARN IN PUBLIC.
01APPLIED AI/02LLM SYSTEMS/03MLOPS/04SOFTWARE ENGINEERING/
01 / FEATURED WRITINGALL ARTICLES
FIELD REPORT19 NOV 2026
Building an LLM inference node that you can observe.
A single-GPU serving stack is more than a model endpoint. Here is what it took to run vLLM, surface GPU metrics, investigate Grafana dashboards and measure short-prompt load.
READ THE CASE STUDY HOMAAI / EVIDENCE REGISTER ↗
MODELGEMMA 4 · 12BSERVINGVLLM · BF16METRICSPROMETHEUS + GRAFANASTATUSFIELD NOTES / NOT AN SLA
1,820requests completed
0observed errors
10mclosed-loop test
OBSERVED SHORT-PROMPT LOAD ONLY · NOT A PRODUCTION SLA. METHOD ↗
02 / SELECTED SYSTEMSTHE FULL INDEX
Not just posts.
Each entry points to a different kind of evidence: an operations write-up, public source code, or a documented workflow.
03 / TWO WAYS TO EXPLORECONTENT, CONNECTED.