You’re in good company

A few of the companies we’ve worked with.

When a good pilot stalls

A model that works in a demo is a promising start. Running it every day takes more: the data it learns from, the checks it passes before each release and a way to know it’s still right. That work is often called MLOps. Without it, a pilot stays a pilot, or goes live and can slowly get worse without anyone noticing.

Two kinds of model, one discipline

Some models you train on your own data, to predict, classify or check an image. Others you build on: a language model from a provider, shaped by your prompts and the documents it can search. The work differs, but what production needs is the same.

  • Models you train

    The work is in the data: collecting it, labelling it, keeping each version and retraining when the world it describes changes.

  • Models you build on

    The work is in what surrounds the model: the prompts, the documents it searches and the checks on its answers, and testing again when the provider releases a new version.

  • Either way

    Every version recorded, every change tested before release, every live model watched, and a quick way back.

Data your models can learn from

A model is only as dependable as the data underneath it. We build the platform that keeps that data in order, so every result can be traced back and every training run repeated.

One result, a delivery estimate that reads Arrives Thursday, traced back: it came from the Delivery estimate model, version 4, which was trained on a recorded set of data, with a recorded version of the code and recorded settings.

  • Data with versions

    Each training run records the exact data, code and settings it used, so it can be repeated and compared.

  • The same inputs, in training and live

    Inputs are prepared once and used in both, so a model sees in production what it learned from.

  • A record of every model

    One place (a model registry) lists each model and prompt: its version, how it was tested, where it runs and who owns it.

Data Engineering
A release check: a new version compared with the live one on the same cases. A typical request, a case that matters: same. A rare one: better. Missing information: same. One it should refuse, a case that matters: worse. So the release waits for review.

Tested before every release

Every change is tested before it reaches anyone: a new model, a retrained one, a new prompt or a provider’s new version. Each is scored on a set of real cases, including the hard ones, and goes live only if it does at least as well where it matters. If it falls short on a case that matters, it waits for a person to decide.

How to test an AI feature before it launches

Released safely, reversed quickly

A new version can run alongside the current one first, answering in the background so the two can be compared (shadow mode). Then it serves a small share of requests (a canary release), then everyone. If something looks wrong, you go back to the last good version, a step that’s tested before it’s needed.

One request, When will my order arrive?, answered by both versions. The live version’s answer, Arrives Thursday, is sent. The new version, running alongside it, answers Arrives Friday: different from the live answer, and not sent. Beneath them, the way back: go back to the last version.

Watched once it’s live

  • Quality

    Samples of real results reviewed by people, alongside feedback from the people using them.

  • Changing data

    An alert when the inputs start to look different from what the model learned from.

  • Speed and cost

    Response times and the cost of each request tracked, so neither grows unnoticed.

  • Uncertain cases

    When the model isn’t sure, the case goes to a person, and their decision helps improve the next version.

Every live model is watched on what matters to your business, and each check has an owner who hears about a problem first.

Kept accurate as things change

Your business changes, and so does the data your models learn from. Providers release new models and retire older ones. We plan for both: retraining when the evidence calls for it, and testing a replacement on your own cases before anything switches.

Then we keep running the platform with you, or hand it over with the code, the test cases, the record of every model and the documentation.

Already have a model that works in a notebook but not in production? We can review it, make it ready to run and look after it from there.

Work with us as a dedicated team, on a fixed price project or a mix of both.

Compare engagement models

Let’s ship it better

Tell us what the model needs to do, where it runs today and what has to be true before you can rely on it. We’ll talk through the options and a sensible first step.

Thanks, we’ll be in touch soon.

{{ ctaStatus }}