Notes on AI that actually helps.

What we learn building AI tools for real businesses — written in plain language, with real examples. No hype, no names.

AI agents & automation

21 July 2026·5 min read

Observability for AI agents: why "it worked in the demo" is not a measurement

A chatbot that handled the ten questions in a demo perfectly can still fail silently on the eleventh, in production, with no one noticing. Observability and structured evaluation are what turn "it seems to work" into a number you can track.

Read →

AI agents & automation

21 July 2026·4 min read

MCP (Model Context Protocol): the plumbing that lets an agent use your company’s tools without a custom build each time

Every tool an agent connects to used to mean a bespoke integration. The Model Context Protocol standardizes that connection — one interface, many tools, no rebuild every time you add a new one.

Read →

AI agents & automation

21 July 2026·5 min read

AI agents for medical secretarial work: what a randomized trial found about time saved and burnout

An AI agent that drafts a consultation note from the conversation is one of the few agent use cases with a real randomized trial behind it. What it found — and what it did not — sets realistic expectations for the sector.

Read →

AI agents & automation

21 July 2026·5 min read

Predictive maintenance agents in industry: reading sensor drift before the breakdown, not after

Predictive maintenance is one of the oldest applications of machine learning in industry — and one of the clearest examples of a mature field where an AI agent adds a layer, rather than reinventing the method.

Read →

Strategy, costs & ROI

21 July 2026·4 min read

The real running cost of an AI agent: it is not the build, it is the tokens

The invoice for building an agent is a one-time number. The cost that determines whether it stays profitable is the one paid per request, forever — and it has been falling fast enough to change what is worth automating.

Read →

Business & compliance

21 July 2026·5 min read

Why an agent should see less data, not more: minimization as a design rule, not a compliance afterthought

The instinct when building an agent is to give it broad access "just in case it needs it." Research on what large models retain from what they process shows why that instinct is exactly backwards.

Read →