https://apeg.dev/ Forward deployed engineering, agentic AI and CPU inference, written from real engagements and measured results. By Arun Prasath E G. en-us Tue, 15 Sep 2026 09:00:00 GMT https://apeg.dev/logo.pnghttps://apeg.dev/ https://apeg.dev/forward-deployed-engineer/how-it-runs/ https://apeg.dev/forward-deployed-engineer/how-it-runs/ Mon, 14 Sep 2026 09:00:00 GMT Arun Prasath E G Forward deployed engineering as an operating model, from the records of real engagements: what is written before day one, what ships every week, how the build stays safe at speed, how credentials and platform rules are handled, what the handover contains, and when the model does not fit. https://apeg.dev/forward-deployed-engineer/ https://apeg.dev/forward-deployed-engineer/ Mon, 14 Sep 2026 09:00:00 GMT Arun Prasath E G A forward deployed engineer builds the AI system inside the customer's business, with the customer's data, under the customer's rules. What the job actually is, from someone who does it. https://apeg.dev/forward-deployed-engineer/for-owners/ https://apeg.dev/forward-deployed-engineer/for-owners/ Mon, 14 Sep 2026 09:00:00 GMT Arun Prasath E G What changes in the first month, what the owner is asked to decide, what the owner is not asked to do, and what the owner's team owns at the end. From the decisions files and handover documents of real engagements. https://apeg.dev/forward-deployed-engineer/retail/ https://apeg.dev/forward-deployed-engineer/retail/ Mon, 14 Sep 2026 09:00:00 GMT Arun Prasath E G What AI in retail looks like from inside a retail chain's operating platform: the data reconciliation that comes before any model, the monthly forecast worked by hand, the five feeds that make it store-aware, the model ladder, open-to-buy, and where AI is the wrong call. https://apeg.dev/forward-deployed-engineer/vs-solutions-engineer/ https://apeg.dev/forward-deployed-engineer/vs-solutions-engineer/ Mon, 14 Sep 2026 09:00:00 GMT Arun Prasath E G The four roles side by side: when each one works, what each one owns, and which one you actually need. Plus how a forward deployed engineer differs from a software engineer and an AI engineer. https://apeg.dev/writing/self-hosted-llm-without-a-gpu/ https://apeg.dev/writing/self-hosted-llm-without-a-gpu/ Fri, 11 Sep 2026 09:00:00 GMT Arun Prasath E G A self-hosted LLM is the answer when your data cannot leave the building. What leaves in a normal deployment, the four ways to keep it in, how to size by workload rather than by model, which open-weight models fit, what it costs to run after day one, and when self-hosting is the wrong call. https://apeg.dev/writing/running-gemma4-26b-on-a-cpu/ https://apeg.dev/writing/running-gemma4-26b-on-a-cpu/ Tue, 30 Jun 2026 09:00:00 GMT Arun Prasath E G A 26B mixture-of-experts model on a desktop CPU with no graphics card: about 40 tokens/sec single-stream lossless, about 124 aggregate when batched. The byte budget, the quantization recipe, the measured memory bandwidth, the dead ends, and the exact commands to reproduce it.