What actually works when you put AI to work on real business systems — the tools worth your money, the ones that fall over in production, and the shortcuts that turn a quarter-long IT project into a week. Written between builds, not by a content team.
By Michael Daugherty — Principal AI-Native IT Engineer
11 notes
Four months of AI coding sessions is 396MB of JSONL nobody can read. I turned it into a searchable record, put a live board over it, and gave the whole thing an agent that can act on what it finds.
A retrieval endpoint can only ever tell you about the past. Swapping it for the Claude Agent SDK took an afternoon and changed what the thing is for.
Semantic search finds what you meant and misses what you typed. Keyword search does the opposite. Reciprocal rank fusion over three lanes in one Postgres query fixed both, and the tuning flags matter more than the embeddings.
The best-sounding open TTS models are almost all non-commercial, and the badge on the repo does not tell you that. Here is the survey I ran before putting a voice in my own tooling.
Local dictation is a solved problem for about four minutes, until you leave the microphone open during a pause and the model invents a sentence. Here is what actually breaks and what fixed it.
The bill was running $2,500–$3,000 on a busy day, and almost all of it was inbound calls sitting in a Twilio IVR. Nobody was buying software. They were renting a phone menu by the second.
I support all four and I'll happily work in yours. But after twenty years of wiring systems together, the honest answer is that a custom integration is now sturdier, more flexible, and often faster to build than assembling the same thing out of connector boxes.
Six policies that stop the attacks small businesses actually get hit with, in the order I deploy them, plus the break-glass account that keeps a bad policy from becoming a very long evening.
Cancelling subscriptions pays for the build. That's the easy math. What makes a consolidation stick is the seams between the systems, and sequencing it so each step cancels a bill before the next one starts.
An honest field report after shipping real internal tools with an AI coding agent: the parts where it saves weeks, and the parts where it quietly creates a mess, and the guardrails that decide which one you get.
Pointing a language model at a stack of PDFs is the easy part. Here's the architecture that keeps a confident wrong number from becoming a paid invoice.
Most of these notes started as a real problem on a real tenant. If one of them sounds like your week, that is usually a short conversation.