Services / AI and machine learning
It is keeping it working in a product, on data that keeps changing, at a cost per request that still makes sense when you have customers. That is the work.
Fixed-fee discovery · Feasibility · Production
The work
Described as work rather than as a field. Any stage can stop the project early, which is the point of doing them in order.
Given your data, what could reasonably be predicted or extracted, and what could not. Usually the most valuable week of the project.
Labelling, cleaning, and finding out what is actually in there. Nearly always bigger than expected.
Choosing an existing model and fitting it to the problem, more often than training something. Faster, and usually better.
Retrieval over your own documents, fine tuning where it earns its cost, and the evaluation harness that separates better from merely different.
Serving, latency, cost per call, and what the user sees when the model is wrong or unavailable.
Models drift as the world moves under them. Something has to watch for that, and it is never the model itself.
Expectations
Said here so they arrive as planning inputs rather than mid-project discoveries.
01 / Usually the biggest line
Teams budget for modelling and discover that most of the hours went into getting data into a state where modelling was possible. It is the least glamorous line in the plan, and frequently the largest.
Line one of the plan, more often than not
02 / Designed, not patched
Confidence is low on some requests no matter how good the system is. What the user sees in those moments is a design decision, and it decides whether the feature is trusted.
03 / Priced before the build
Affordable in testing can be ruinous at customer volume. The estimate belongs in discovery, because occasionally the number kills the idea before the build wastes it.
How the work runs
Everything else on this site follows from these three steps.
A short piece of paid work. We sit with the problem and you get back a document that says what is being made, how it fits together, what order it gets built in, and what it costs. The plan is yours unconditionally. Take it to us, take it to another firm, or decide not to build at all.
The people who understood the problem build the thing. You get names rather than a rotating cast, and one lead who answers to you directly. If someone is not working out, we raise it first and manage the replacement ourselves.
Work arrives in increments you can see and use, not in a reveal at the end. When scope moves, and it will, the change is costed in writing before anybody starts it.
Change orders
Priced in writing, before work moves. Scope changes on real projects; what never changes silently is the bill.
Where discovery differs here
Discovery on machine learning work often ends with a feasibility answer instead of a build plan. Sometimes the recommendation is not to proceed, and the document says so plainly. That outcome costs a fraction of a failed build, so it counts as the engagement working.
Ownership
From the first commit, everything belongs to you. Not at handover. From the beginning.
All of it, in accounts you control. There is no private framework underneath and no dependency on continued payment to keep the thing running.
Cloud accounts, domains, credentials and certificates are created in your name. We hold nothing that you cannot revoke yourself.
Decisions, diagrams and runbooks are written while the work happens, so the knowledge leaves with the documents rather than with the people.
If you stopped paying us tomorrow, the product would keep running without us. That is the test we design against.
Questions
Next step
Say what you hoped a model would do and what data you have. If the honest answer is that you do not need one, the reply will say so.
A few lines about the situation is plenty. We read every one.
Expect a reply within two business days. If we are not the right fit for the work, the reply will say so rather than dodge.