Notes from the build.
Engineering deep dives, architecture decisions and what we learned shipping operations-heavy software.
AI1 Oct 2026 · 4 min read
Prompt Injection Explained for Product Teams
Prompt injection lets a stranger steer your AI feature through an email or web page. Why it happens, what is at risk and the design choices that limit the damage.
Evaluating an AI Feature Before Your Users Do
A good demo says little about production. Learn how to define quality, build a test set from real work and catch AI failures before your users find them.
Retrieval Before Fine Tuning
Teams reach for fine tuning to teach a model their documents. Retrieval is usually cheaper, easier to update and easier to audit. Here is when each approach belongs.
Where Large Language Models Earn Their Keep in Operations
Language models are strong at reading messy text and weak at exact decisions. We map where they pay off in operations, where they disappoint, and how to design for their mistakes.
Open Weights or Hosted APIs: A Practical Comparison
Hosted APIs are fast to start and open weights give control, but the real costs sit elsewhere. A practical way to choose, test and keep your options open.
Small Models, Big Wins
A small model trained on your own data can beat a large one on speed, cost and predictability. When to choose one, and when not to.
Human in the Loop Is a Design, Not a Disclaimer
Saying a person reviews everything is not a control. Reviewers tire, queues win, and approvals become clicks. What it takes to design human oversight that works.
Structured Output Changed How We Build With LLMs
Moving from regexes over free text to schemas changed how we treat models: as typed components we can test, validate and replace. Here is what it fixed and what it does not.
Why AI Pilots Stall at the Handoff
The demo worked, the model was accurate, and the pilot still never shipped. The failure usually sits where AI output has to become someone's real work.
The Real Cost of an AI Agent
The cost of an AI agent is rarely the model bill. Review time, integrations, evaluation and exceptions add up, and a business case that ignores them will not survive contact with real volume.
Have a problem worth solving?
Tell us what is not working, what you are trying to build, or where you want to go next.