ML Env

How we work

Audit first. Build second.

From your first email to a system your team relies on is usually two months, and you can stop at the end of any phase with everything produced so far. Below is what happens in each, what you get at the end of it, and what we need from you to start it.

  1. 00

    First conversation

    One week. No charge.

    You describe the process that eats the most time. We tell you whether a language model is the right tool for it — and say so plainly when it is not, which happens more often than the industry admits.

    What you get
    A straight answer, and a rough sense of the work involved.
    What we need
    A description of the process in your own words. No brief, no NDA yet.
  2. 01

    Audit

    About two weeks

    We map your workflows and rank them by what a model would actually improve. We check that the data each case needs already exists, because most cases die there rather than in the modelling. Some engagements end at this phase, and that is a fine outcome.

    What you get
    A written report: a ranked shortlist, the data each case needs, the running cost of each on our hardware, and an honest list of the cases we would not build.
    What we need
    A few hours from whoever does the work today, and read access to the systems that hold the data.
  3. 02

    Prototype

    Two to three weeks

    We stand up the highest-value case on our hardware and put it in front of your team with your real data. You judge it on your own examples, not a demo. If it fails on them, we would rather find out now than after the integration.

    What you get
    A working prototype and an evaluation set built from your examples, so quality is something you measure rather than feel.
    What we need
    Your real data, and a few people willing to be sceptical about the output.
  4. 03

    Deploy

    Two to four weeks

    The system runs on our servers or yours, whichever fits your compliance and budget. It goes in behind a rollback path, with a person in the loop wherever a wrong answer would reach a customer or a ledger.

    What you get
    The case running in production, integrated with your stack, monitored. A fixed monthly fee, with no per-token meter.
    What we need
    Someone on your side who can approve a change to the shop.
  5. 04

    Hand over

    Ongoing. Leave with 30 days' notice.

    We watch it, patch it, and move it to a better open model when one arrives. Nothing about the system is hostage to us: the weights are open, the prompts are yours, and the runbook says how to run it without us.

    What you get
    The weights, the prompts, the evaluation set and the runbook — at the end, and at any point before it.
    What we need
    Nothing. That is the point of this phase.

Where the timeline slips.

Almost always on data rather than on modelling. An export that turns out to be a print-layout PDF, discount rules that live in the owner's head and were never written down, or a catalogue whose product records disagree with the warehouse. The audit exists to find these before you have paid for a build, which is why we will not skip it — including when a customer asks us to.

Next

Tell us what your team does by hand

Describe the process that takes the most time. We will say plainly whether a model is the right tool for it.