Loading articles
Practical AI engineering: prompt optimization, agent workflows, and integrating LLMs into production applications.
10 posts in this category
Last time, Laya learned my classifiers from Jev through a paid API. This time there was no API key: Claude Code and Codex, on personal subscriptions, wrote 2,000 training inputs each and labelled every one. Laya went from 31 to 40 of 50 test cases with Opus 5.5 as the teacher and to 38 with Sol 6.1, two and four cases behind the Jev-taught model. Codex on a $20 plan ran out of its five-hour window before it finished.
In the last two posts, untrained Laya lost every race. So I fine-tuned it on my six-year-old desktop's RTX 3050: once on my Pigeonhole classifiers, with Jev as the teacher, and once on Pac-Man, with a twenty-line expert as the teacher. It went from 31 to 44 of 50 test cases, level with Cloudflare's Clef, and from zero wins to 87 out of 200 games. Total bill: under thirty cents.
Two weeks after I raced Jev against Laya, Cloudflare released Clef and OpenAI released a Decisions API. I ran all of them on the same 74 test cases through Pigeonhole. GPT-6 Luna nearly tied Jev and was the fastest hosted model, Clef-flash was close behind at half the price, and one refused question taught me something about batching.
Meta has Muse, OpenAI has Dots. I wanted an AI pet that animates itself live with no image generation, works with any AI provider, and plugs into any API. The trick: let the model pick, and let a renderer draw.
Pigeonhole is a weekend project that turns a plain-English description into a tested classification API. I ran it on Jev (hosted, a fraction of a cent) and an out-of-the-box, untrained Laya (local, free) side by side on my own machine. Here's what each got right, how fast they were, and what it takes to run Laya yourself.
I got tired of agents handing me giant Markdown blobs I would never read properly. So I built MaraDocs to turn agent reports into durable HTML artifacts.
Practical AI engineering — not hype. These articles cover what it actually takes to build with large language models in production: designing agent workflows that plan and execute real work, structuring prompts that stay reliable as requirements change, and fine-tuning models when off-the-shelf behavior isn't enough.
Everything here is written from hands-on experience shipping AI features, with runnable examples and the trade-offs spelled out.