Vilozi Engineering
· 6 min read
There is no best AI model. There is a best model for each task, and it changes every few months. Vilozi Intelligence, the layer under our platform, exists to make that a routing decision instead of a rewrite.
Different jobs, different strengths
Our platform asks AI to do very different things. Understanding a caller in real time needs speed. Turning a long job description into line items needs careful reasoning. Reading thousands of records needs low cost per item. Frontier models from Anthropic, OpenAI and Google each lead on some of these and trail on others, and our own smaller models, tuned on industry work, beat all of them on a few narrow tasks.
Evaluate on real work
Public benchmarks don't tell us which model will write a better roofing quote. So every task has its own evaluation set built from real, permitted job data, with the answers checked by people. A new model, or a new version of an old one, runs against that set before it gets any live work. If it's better on quality, speed or cost without being worse on the others, it takes over that task. If not, nothing changes.
Canaries before switches
Even a model that wins the evaluation goes live gradually. It first handles a small share of real traffic, its answers compared against the model it would replace. Only when the two agree, or the new one is measurably better, does the switch finish. A switch is a configuration change, not new code, so it can be undone in minutes.
Where our own models fit
General-purpose models are trained to be good at everything. Many of our tasks are narrow and repeat thousands of times: classifying a job, pulling a scope out of a description, matching a request to the right crew. For those, a smaller model tuned on industry work can be faster, cheaper and more accurate than a frontier one. We build these where the evaluation says they win, and nowhere else.
Privacy by design
Outside AI providers never train on our customers' data, and our own models are tuned only on data we have permission to use. Routing gives us the freedom to choose the right model for each task without giving up control of the data behind it.
Vilozi Engineering
Notes from the Vilozi engineering team on the open problems we work on.
