Models and AI engineering
It has read the internet and none of your tickets. We take an open-weight model, train it on the work your team has already done, and measure it against the one you pay per token for today. When ours wins, it runs in your own infrastructure, costs a fraction per request, and the weights are yours.
What we train
We start from the task, not from the technique. Sometimes the honest answer is that you do not need a trained model at all, and we would rather say so before anyone signs anything.
We take an open-weight model, keep training it on the work your company has already done, and end up with one that answers and reasons the way your team does. It runs on your hardware and the weights are yours.
Most of what companies ask a large model to do is one narrow job repeated a million times: sort this, pull that field, score this call. A small model trained for that one job is faster, cheaper, and far easier to defend in an audit.
When the right answer is a matter of judgment rather than a label, we train on the judgment itself: what your reviewers accepted, what they rewrote, and what they threw out.
The part that decides whether any of the above survives contact with production: a test set that catches regressions, a registry that says which version is live, and serving that holds up under your real traffic.
Before you train anything
Every step down this list costs more than the one above it and takes longer to change your mind about. We work down it in order, and most engagements stop before the bottom.
We charge for the evaluation that answers this question, and we charge for it separately, so hearing “you do not need to train anything” costs you two weeks instead of a project.
Where it pays off
None of these need a model that can write poetry. They need one that is right about your work, fast, and cheap enough to run on everything that comes in.
Every message sorted by team, urgency, and language before a person opens it, trained on ten years of how your team actually sorted them.
Invoices, delivery notes, and contracts turned into fields your systems can read, including the scanned ones and the note somebody wrote in the margin.
A model trained on your own images or sensor history that flags the bad batch on the line instead of at the customer.
Trained on your sales, your seasons, and your disruptions, rather than on a generic curve that has never seen your market.
Every call scored against the criteria your own supervisors use, so coaching covers all of them instead of the three someone had time to listen to.
A first pass over documents against your internal rules, with the clause it tripped on attached, so your reviewers start from a shortlist.
How we engineer it
A notebook that scored well once is not a deliverable. These are the things we build alongside the model, and together they are usually more work than the training itself.
We write the test set before we train anything. If we cannot measure the task, we cannot tell you whether we improved it.
Every run is pinned: data version, code version, seed, and hardware. A result nobody can reproduce is a rumor.
We record where every training example came from and what you are allowed to do with it, before it ever reaches a GPU.
Models are versioned like code. You can see which one answered a given request, and go back to the previous one in minutes.
A target for price per request and response time is set at the start. A model that misses it is not finished, however good the accuracy looks.
Your business changes and the model quietly stops being right about it. We watch for that and tell you when it is time to retrain.
How we work
The first phase exists to talk you out of it. It is cheaper for both of us to find out in two weeks that a task is not learnable than in four months.
01Estimate: 2 weeks
We define the task precisely, build the test set, and measure what an off-the-shelf model already scores on it. That number is what everything afterwards has to beat.
DeliverableA test set, a baseline score, and a recommendation
02Estimate: 2–4 weeks
We assemble the training set out of work your team has already done, clean it, label what needs labeling, and hold back the hard cases for the exam.
DeliverableA versioned training and evaluation set
03Estimate: 3–6 weeks
We run the experiments, compare sizes and techniques, and stop at the smallest model that clears the bar you set.
DeliverableTrained weights and the runs that produced them
04Estimate: 1–2 weeks
We score it against the baseline and put its worst outputs in front of the people who do the job today, because a benchmark never catches everything.
DeliverableAn evaluation report and a go or no-go
05Estimate: 2–3 weeks
We serve it inside your infrastructure behind an API your systems can call, with a fallback for the cases where it is not sure.
DeliverableA served model with monitoring and runbooks
06Estimate: Ongoing
New examples pile up, quality drifts, and the model gets retrained on a schedule your team can run without us in the room.
DeliverableA retraining pipeline and the handover
The timings on each stage are a reference from past projects.
Stack
We use what you already pay for wherever it holds up. When there is no constraint, this is where we start.
How we engage
Most companies start at the top and decide after it, with a number in hand instead of a hunch.
Two to three weeks, fixed price, to answer one question: can a model do this well enough to be worth building, and what would it cost to run?
Best for
Teams with a task in mind and no idea whether it is learnable.
Includes
End to end: data, training, evaluation, and a served model your systems can call, with your team in the room for all of it.
Best for
A task that cleared feasibility and has a budget behind it.
Includes
A monthly retainer for the part nobody plans for: the model is live, the world moved, and something needs retraining.
Best for
Teams running models in production without an ML engineer on staff.
Includes
Your model
Everything we train runs in your accounts, on base models whose licenses let you keep what comes out. When the engagement ends you hold the weights, the training data, and the code that produced both, and you have no technical reason to call us again.
Questions
Often you do not, and the feasibility study is how we prove it either way. If a prompt and your own documents get you to 90% of the value, that is the answer we will give you, and it costs two weeks to find out instead of a project.
For a classifier or an extractor, a few hundred good examples is often enough to beat a general model. For fine-tuning a language model to reason the way your team does, a few thousand. Quality matters far more than count, and “we have ten years of tickets” is usually a better starting point than people expect.
It is. Everyone's is. Cleaning and labeling is a real phase with real weeks attached, which is why it sits on the plan above instead of hiding inside the training step. The one thing we cannot fix is data that never recorded the thing you want to predict.
A feasibility study is fixed price and a small fraction of a build. A full build depends on the task, and training is rarely the expensive part: data and evaluation usually are. Running it is the easier number, because a small task model on your own hardware is generally much cheaper per request than an API call, and we project that for you before you commit.
For plenty of things you should, and we build those systems too. Training earns its place when volume makes per-token pricing hurt, when latency matters, when the data cannot leave your building, or when the task is narrow enough that a small model beats a large one at it. If none of those apply, we will point you at an API and save you the project.
For training, no. We rent them for the weeks we need them and pass the cost through. For serving it depends on the model: plenty of task models run fine on CPUs you already have, and the ones that need a GPU usually need less of one than people assume.
Describe one repetitive judgment call somebody on your team makes all day. We will tell you whether a model could learn it, what it would take to find out properly, and what it would cost to run once it works.
The first conversation is free. If a prompt would solve it, we will tell you that instead of selling you a training run.