How we work
We build in our lab. Your data stays with you.
NPU Labs is a specialist lab that companies bring in for AI engineering. The work happens at our Cape Town office, where all our models, NPUs, GPUs, and other hardware are within reach. Experiments run on synthetic data, public datasets, and our own test scenarios, never on yours.
Our lab
Cape Town office
- Every model we run
- NPUs, GPUs, and other hardware
- Synthetic data, public datasets, our own scenarios
Your environment
Your premises or cloud
- Your data
- Your hardware or private cloud, your access controls
- Run by our engineers or your team, your choice
Why our lab
Everything the work needs is already in the room.
We work from our office because that is where the models, the hardware, and the test sets we have built up all are. You get the benefit of all of it without standing any of it up yourself.
Every model we run, side by side
Candidates are tested against each other across the models in our registry, not limited to whatever one environment happens to have installed.
See the model registryNPUs, GPUs, and other hardware
A workload is measured on the kind of hardware it will run on before anything is specified, so a recommendation comes from a benchmark, not a guess.
See the hardwareNothing to provision first
You do not have to set up compute, accounts, or access to your data before we can start experimenting. The lab is already running.
Your data
Experiments never touch your data.
Nothing we try out in the lab runs on your records. When the system finally has to work with your data, that step happens in your environment, and you decide who runs it.
What our experiments run on
Synthetic data
Generated to match the shape of your problem: the formats, the fields, and the edge cases, with no real records in it.
Public datasets
Open datasets and published benchmarks.
Our own scenarios
Golden scenarios and test sets we have built up across our own products.
When it meets your data, you choose who runs it
Option A
We run it in your environment
The steps that need your data, such as fine-tuning, indexing documents for search, and evaluating on real cases, run on your hardware or private cloud, under your access controls. Our engineers operate them there, and your data does not come to our office.
Option B
Your team runs it
We hand over the pipelines and runbooks, and your own people run every step that touches your data. Our engineers never see it.
Either way, the same rule holds: your data stays in your environment.
Where the work happens
In person where it counts. Remote for the rest.
Every step of an engagement has a place. We come to you for the kickoff and the install, the build happens in our lab, and anything that needs your data happens in your environment.
- 01
Kickoff
At your offices
Discovery workshops with your team: the systems involved, the constraints, and what success looks like.
- 02
Build and experiment
In our lab
Models, pipelines, and evaluations built and tested on our hardware, on synthetic data, public datasets, and our own scenarios.
- 03
Meet your data
In your environment
Fine-tuning, indexing, and evaluation on real cases, run by our engineers inside your environment or by your own team.
- 04
Install and hand over
On your premises
When hardware is part of the build, we install it on site and hand it over with its runbook.
Everything in between is remote. Check-ins, reviews, and demos run over video call.
Ways to engage
Four ways to work with us.
Pick the one that matches where you are. If you are not sure yet, start with an assessment.
Short, fixed scope
Assessment
A focused look at your systems, costs, and constraints. You leave with a recommendation on where AI earns its place and what to build first.
Ask about an assessmentA defined project
Fixed-scope build
A defined deliverable, built and tested in our lab, brought to your data in your environment, and handed over with its runbook.
Scope a buildAfter handover
Managed service
We keep it healthy on an ongoing arrangement, and the engineers who built it stay on it.
Ask about managed serviceMonth to month
R&D retainer
A standing arrangement where we run experiments and builds for you in our lab, month to month, under the same data rules.
Ask about a retainerCommon questions
Questions we get asked about working with us.
Do you need access to our data to get started?
No. Experiments run in our lab on synthetic data, public datasets, and our own test scenarios. Your data only comes into it at the step where the system has to work with it, and that step runs in your environment.
Will our data be used in your experiments or for other clients?
No. Experiments never use client data, and your data never comes to our office. It stays in your environment, on your hardware or private cloud, under your access controls.
Who runs the steps that do need our data?
You choose. Our engineers can run them inside your environment, under your access controls, or we hand your team the pipelines and runbooks and your own people run them.
Do your engineers work on site?
For the kickoff workshops and for hardware installs and handover, yes. The build and the experiments happen in our lab, and check-ins, reviews, and demos run over video call.
Which engagement should we start with?
If you are not sure yet, start with an assessment. It is short and fixed in scope, and it ends with a recommendation on what to build first and which way of working with us fits it.
Not the question you came with?
Ask us about your projectNot sure which engagement fits?
Tell us what you are trying to build. We will tell you which way of working fits it, what it takes, and whether we are the right people to build it.