Hire AI and ML developers
Add a developer for the machine learning side of your product: what model, over what data, measured how, and at what cost per run.
The retrieval, prompting and evaluation in our own products are built and measured by this team, with the spend recorded per call.
Where it fits
Search over your own documents
Retrieval that returns the right passage, with the sources shown and the failures counted.
Choosing a model
A comparison on your own cases, with quality, latency and cost per run side by side.
Classification and extraction
Turning messy text into fields your systems can use, with a measured error rate.
The data work first
Collecting, cleaning and joining the data a model needs, within each source's terms.
What is in it
- What they work in
- Python, PyTorch where training is involved, embeddings and vector search, and the hosted model APIs you already use.
- A measurement before an opinion
- Claims about quality come with the test set they were measured on and the date.
- Cost per run, written down
- What a request costs at your volume, before it goes live, so the bill is not the surprise.
- A named person, introduced first
- You meet them and see what they have built before anyone starts.
- A monthly engagement
- Engaged by the month at a rate agreed before they start, with a month's notice either way. No hourly billing.
What is out
- Permanent placement and recruitment
- Training a large model from scratch
- Scraping data a source's terms forbid
Questions, answered
Do we need machine learning, or just an API call?
Most products need a hosted model and good retrieval, not training. We will tell you which one your case is on the first call, because the cheaper answer is usually the right one and it is easy to check.
Do you train your own models?
We fine-tune where there is a measured reason to, and we do not train large models from scratch. If your problem genuinely needs that, we are the wrong firm and will say so.
How do you keep model costs under control?
By knowing the cost per run before launch and watching it after. Model spend is billed to your own provider account at cost, so you see the real number rather than ours.
Can they work with our data team?
Yes. They take work from your backlog and sit in your process. Where your data lives in systems we should not see, we work on anonymised or synthetic copies and write down any exception.
Tell us what you are building.
Thirty minutes, free, with the people who would do the work.
Book a call