AI, agents & language models.
AI that does a job, rather than a chat box bolted to the corner of a page. Built into the thing it is meant to help, and grounded in your own material so it cannot invent an answer.
Assistants grounded in your own content
A helper that answers from your prices, your documents and your policies — not from whatever a general model absorbed about your industry. Retrieval runs first: the question is matched against your material, and only what matches is put in front of the model.
The part that matters is what happens when there is no match. A question your content cannot answer is refused rather than guessed at. A model that confidently invents a price you do not charge is worse than no assistant at all, because somebody acts on it. The helper on this site works exactly this way: it quotes real numbers from the same source the pricing page reads, and says plainly when it has nothing to go on.
Self-hosted models, on hardware you own
Per-request billing grows with your traffic and sends your input to somebody else’s server. Neither is acceptable when the input is client records, contracts, or anything you would not paste into a public tool.
So the model runs on a machine you own. This site’s own helper runs on a GPU in a house, and the engineering around it is the real work: embeddings computed ahead of time so a question costs one short lookup instead of a long generation, rate limits per visitor and a ceiling for the whole day so an open endpoint cannot run up a bill, and a health check that removes the feature from the page when the machine is off — rather than leaving a button that accepts a question, thinks for twelve seconds and apologises.
Automation that fires on real events
“When this happens, do that” — a booking confirmed, a payment cleared, an enquiry left sitting for three days. Run inside the request that caused it, with three guards that all have to hold before anything is sent: each rule fires at most once per record, there is a daily ceiling that fails closed if it cannot be read, and an action can never trigger another action. The worst realistic failure here is mailing a client repeatedly, and that cannot be taken back.
Machine learning and computer vision
Where a language model is the wrong tool. Data cleanup and analysis, scripts that remove repetitive work, and vision projects built for the problem in front of them — including real-time sign-language recognition running on a Raspberry Pi. Practical models and pipelines rather than buzzwords.
What it costs
Quoted rather than priced from a list, because no two are the same shape and most of the cost is in the plumbing rather than the model. Scoping and advice run at $120 an hour, build work at $85 an hour, and the quote comes before anything starts. Self-hosting has a hardware cost that is yours to keep; renting inference does not, and which is cheaper depends entirely on volume — I will tell you honestly which way your numbers fall.