Notes from the workshop.
What actually worked, what it cost, and where it broke — written up while the details are still fresh.
Tagged Testing · show all_
An AI agent on cheap hardware: the project, the process, the numbers
A full build log for a local email-triage agent: the target it had to hit, three model builds, eight prompt versions, two rejected releases, and the scores at each step.
read_Update: the local AI agent is three times faster, and becoming a mail app
Same desktop, three times the throughput: about 17,000 emails a day, 41 of 45 real emails filed right, zero legitimate mail binned — and the triage engine is turning into a full mail app.
read_Update: the local AI agent is now a mail app. The mail stays on the machine.
The triage engine is now a mail app for the web, iPhone and Android: one inbox over every account, mail filed by client and job, AI drafts, dates out of a photo. The mail is read on the machine. The gate held back two prompts on the way and passed a third.
read_Your AI reads your chats. Then it calls the police.
Claude and ChatGPT train on consumer chats by default. Researchers pulled training data back out of a live model for a few hundred dollars. Claude's share button put medical notes and children's phone numbers on Google. Loi 25 calls that paste a communication of personal information outside Québec.
read_High availability or good backups?
High availability keeps you online; backups get your data back. Real managed offers for a custom app, from $250 to $720 CAD a month once someone keeps it updated — and how to choose from what an hour offline really costs you.
read_The era of custom software
Shared platforms leak thousands of customers at once. Why a one-of-a-kind app shrinks your blast radius and your attack surface, and why "custom is hard to maintain" no longer holds.
read_An AI employee needs tools: how to give it the right ones
What an AI employee is, why a login, a browser robot or an automation key are the wrong tools, and how MCP gives it exactly the access its job needs — secure, audited, reversible and customised.
read_AI memory, RAG or database: which one does your AI need?
A database answers exactly, RAG answers from your documents, AI memory remembers how you work. A plain guide to choosing — and combining — the three.
read_Vibe coding and the poisoned package
Axios, LiteLLM, chalk, TanStack: the packages vibe-coded apps pull in by the hundred are now a favourite way in. Why "it works" is not the same as "it is safe", and what to do about it.
read_Your data, a chatbot, and no hallucinations
How a document chatbot stays on your documents: four rules, what happens when a question is off the map, how accuracy is measured on your own content, and what the failures taught.
read_Describe your business, get the admin, the API and the MCP
How one schema file becomes a database, an admin, a REST API, access rules and an MCP server for AI assistants — with the checks around it, and what stays hand-written.
read_A “For You” feed with zero AI calls
How a small marketplace built a personalised feed from one signal, an average and a one-in-eight surprise — with no AI model called when the feed loads.
read_Search that understands what people mean: better search with RAG
How a video classifieds marketplace went from zero results to searches that understand intent: the four steps, the numbers, and why showing people what was understood matters as much as the ranking.
read_