Usman Shahid - WritingTechnical writeups and opinions on local LLMs, agent orchestration, and small-model agents.https://codemug.github.io/An SRE's guide to deploying Large Language Models, Part 2: from a base model to a served onehttps://codemug.github.io/blog/sre-guide-deploying-llms-part-2/https://codemug.github.io/blog/sre-guide-deploying-llms-part-2/How a raw next-token predictor is turned into a model that follows instructions, reasons, and stays in bounds, and the GPU, memory, and quantization mechanics that decide whether serving it is fast or ruinously slow.Sun, 02 Aug 2026 00:00:00 GMTllmssrefine-tuningrlhfinferencequantizationAn SRE's guide to deploying Large Language Models, Part 1: understand the workloadhttps://codemug.github.io/blog/sre-guide-deploying-llms-part-1/https://codemug.github.io/blog/sre-guide-deploying-llms-part-1/Before you can deploy and serve LLMs reliably, you have to understand them as a workload. A ground-up tour of the transformer, from tokens to attention to the feed-forward step, for SREs.Sat, 01 Aug 2026 00:00:00 GMTllmssretransformersinferenceI created Primer, a context optimized, loop engineering AI platform so I can get meaningful work done on my 16GB VRAMhttps://codemug.github.io/blog/gaming-pc-agent-platform/https://codemug.github.io/blog/gaming-pc-agent-platform/I bought a 16GB gaming GPU, tried to get real work out of small local models, and ended up building Primer: a platform for running lots of tiny, focused agents instead of prompting one big one.Wed, 29 Jul 2026 00:00:00 GMTlocal-llmsagentsprimercontext-engineering