Long-form pieces on production AI and the engineering work around it.
You can expect deep dives into:
For a shorter professional overview, start with About. For something lighter, see Fiction or Links.
Enterprise teams still lack a standard workflow for adapting, evaluating, versioning, and replacing LLMs on their own terms.
Six lessons from building and shipping enterprise analytics products that use large language models.
Structured context beat naive long context. Retrieval handled noise better. Simple baselines held up better than expected.
Why human review should be an engineering control, not a default rule for every AI system.
Why million token windows do not remove the need for retrieval, structure, and disciplined context design.
Most AI agents still feel like demos. Coding agents are the exception. This is how I use Cursor, Codex, Gemini CLI, Jules, and related tools.
AI makes code cheaper to generate, which raises the value of design, review, testing, and system judgment.
A short first post for the site.