Short-form writing on LLM integration patterns, agent orchestration in production, .NET architecture, and how AI-native tooling changes day-to-day engineering. No hype, no takes — just what's working on my keyboard.
Most LLM-powered analytics look magical in a demo and break in production. A small, boring deterministic compute layer between the model and your UI is what turns a toy into a product. Here's the pattern I've been using in ASP.NET Core services.
Before anyone writes a prompt, I run a 90-minute audit of where staff actually spend their hours. The same five patterns keep coming up — document intake, ticket triage, status reporting, form filling, and knowledge lookup. Here's the checklist I use.
After a year of building agents for real business workflows, one takeaway keeps winning: stop making the model guess, give it a small, well-typed tool surface. Notes on tool design, error boundaries, and the retry loops that actually converge.
A field report from a real migration — breaking a monolith into service-based components with DI, keeping the team shipping features in parallel, and using AI-assisted refactoring where it pays off (and where it absolutely does not).