Skip to content

Your own AI model, developed for your task.

An AI model developed specifically for your task that runs entirely on your own hardware: on a phone, on a laptop or on a server with no internet connection. For work where sensitive data must never leave the device.

  • Developed for one specific task
  • Runs entirely locally, no cloud
  • Phone, laptop or air-gapped server
  • Runs on Qualcomm Snapdragon Apple silicon NVIDIA Intel Arm Android
  • Works with PyTorch ONNX Hugging Face

At a glance

Your own AI model, on your own hardware. Your sensitive data never leaves the building.

  1. Refine.

    Refined for your task, together with your team.

    More about the model
  2. Train.

    On the right data: your own or licensed.

    More about the training data
  3. Test.

    Live only once it passes the test.

    More about evaluation
  4. Run.

    Encrypted on your own hardware, with no cloud.

    More about security

Your own model starts with your idea.

Talk your idea through with an engineer

Local vs cloud

Sensitive data does not belong in the cloud. A local model keeps it all on your own hardware.

With a cloud model every prompt, document and recording leaves the building for someone else’s servers. A local model needs no connection: the work happens on the device itself.

  • Cloud model, over the internet : 1.2 s
  • Local model on the NPU : 35 ms
Example: response time of one short prompt

The answer comes from the device itself.

Nothing goes out: no upload, no data centre and no download.

  • Data stays on the device Not a byte goes to a cloud server
  • Works offline Even with no internet connection
  • No cost per prompt However often your model runs

Where it runs

On a phone, a laptop or a server with no internet.

The secrecy decides the place: the phone for what must not leave the handset, the laptop for a case file that stays offline, and an air-gapped server for what the whole organisation keeps secret.

Runs on

  • Apple silicon
  • Android
Ask about an air-gapped AI server
  • On the phone

    A confidential board meeting is transcribed and summarised on the phone itself.

  • On the laptop

    A lawyer analyses a confidential case file offline, on their own laptop.

  • On an air-gapped server

    The whole company queries its confidential archive on a server that never touches the internet.

The model

Your model starts from the newest open-weight models.

Every new release is tested on your task. The best base is then refined together with your team into a model that knows your field better and works faster than a generic one.

Works with

  • Hugging Face

Choose and refine

Refined for the task, together with your team.

First we agree with you what the model must do and when it is good enough. Then we pick the base that suits the task and refine it with fine-tuning and distillation until it has mastered your terminology and way of working.

  • Decoder SLMs
  • Encoders
  • Vision-language
  • Speech

Every licence is checked up front: the model and weights are yours.

Training data

The right training data for each task.

What a model learns from depends on the task: your own documents, licensed or public datasets, carefully generated examples, or a mix. Everything is labelled by fixed rules, and part of it is kept apart for testing.

Your training data

Put together per task

  • Own, licensed or public
  • Generated where needed
  • A test share is kept apart
  1. Choose

    The sources per task: own documents, licensed or public sets.

  2. Clean

    Duplicates out, personal data masked.

  3. Label

    By fixed rules, so every example is judged the same way.

  4. Generate

    Generated examples where real data is scarce or too sensitive.

  5. Hold out

    Part of it stays out of training, for the final test.

Distillation

A large model teaches your specialist the task.

That makes your model small enough for a phone or laptop while it stays almost as accurate on your task, thanks to SFT, LoRA and DPO.

  • Size

    100×

    Up to a hundred times smaller than the large model

  • Accuracy

    99%

    Almost as accurate, measured on your task

  • Computing power

    1%

    A fraction of what the large model needs

How your specialist learns:

  • SFT
  • LoRA
  • QLoRA
  • DPO
  • Knowledge distillation

Optimisation

From seconds to milliseconds, on the chip you already have.

Your model goes from fp32 to int4 and from gigabytes to megabytes, compiled for the NPU. After every step we check that the answers stay the same.

Works with

Fewer bits, the same answers.

Your model gets fewer bits per weight, the parts it does not need are cut, and it is compiled into one graph for the chip. That way it fits in the memory of a phone or laptop and answers in milliseconds.

Techniques:

  • int8
  • int4
  • AWQ
  • Pruning
  • Operator fusion
  • AOT compilation
meeting-notes 1.7B to 210M parameters NPU · int4
p95 latency per request 38ms
  1. fp32 · baseline 2,140 ms
  2. distilled · fp16 410 ms
  3. int8 · CPU 96 ms
  4. int4 · NPU 38 ms
Memory
7.2 GB to 460 MB
Tokens/s
286
Parity
±0.004

Evaluation

Your model only goes live once it passes the test.

Tested on data the model has never seen: accuracy on your task, your field’s jargon, hard cases and attempts to trick it or draw data out of it. You set the bar together with us up front.

Ready for production

Every threshold met on the test set

  • Accuracy met
  • Jargon understood
  • Hard cases passed
  • Leaks no data

Security

Your model is encrypted and opens only on trusted devices.

On phones and laptops the model is stored encrypted and opens only in a sealed-off part of a trusted device’s chip. The air-gapped server has no connection to the outside world at all.

Your model protected

Sealed off, on every device

  • Weights encrypted
  • Opens only on trusted devices
  • Server with no internet connection

Approach

From feasibility to a model in production:

Start small with a feasibility check, or leave the whole track to us. Training happens where your data is allowed to be: on your own hardware or in a sealed environment without internet access.

The track

Every step is also available on its own. We do the work.

Step 1 of 5 · Feasibility

You see up front what a model of your own achieves.

Together we define the task, how strict the secrecy is and where the model must run. Then we test the best open-weight models on examples of that task, so you see in black and white where you start.

If the data may leave the building, on-⁠demand GPU capacity:

  • B200
  • H200
  • H100

Questions

Questions about AI Model Engineering.

What boards and IT teams ask us most about an AI model of their own that runs locally.

What is AI Model Engineering?

Together with your organisation we develop an AI model for one specific task and run it entirely locally: on a phone, on a laptop or on a server with no internet connection. Not a generic model that does a bit of everything, but a model we choose, refine and test until it does your task better and faster. It is meant for work where sensitive data must not leave the device or the building. The technical term is edge AI: the model runs where the data is.

Why not just use a large cloud model?

Because your data would leave the building. With a cloud model every prompt, document and recording travels to someone else’s servers. For a board meeting, a patient record or the papers of an acquisition, that is not an option. A model of your own runs locally, works offline and costs nothing per prompt. And because it is refined for one task, it often does that task better than a large generic model.

Where does the model run?

That depends on how strict the secrecy is. On the phone for what must not leave the handset, such as a confidential conversation transcribed on the spot. On the laptop for analysis that has to stay offline, such as searching a litigation file. And on an air-gapped AI server in your own building for what the whole organisation must keep secret, such as an archive of research or patent files. Technically it runs on the NPU of modern phones and laptops (Qualcomm Snapdragon, Apple silicon, Intel), on NVIDIA GPUs and on Arm and x86 processors. We compile it for exactly that chip and measure its speed on the real device.

Which base models do you use?

The base is a state-of-the-art open-weight model: decoder SLMs for text, encoders for classification and search, and vision or speech models where the task calls for them. For each task we choose the base on quality, size and licence, then refine it together with your team until it knows your terminology and way of working.

How small can a model get without losing quality?

That depends on the task, and we measure it rather than promise it. A well-defined task, such as transcribing and summarising a meeting or finding unusual clauses in contracts, often fits in a model of a few hundred million parameters, distilled from a large teacher model and quantised to int8 or int4. After every step we test on held-out data: if quality drops below the agreed threshold, that step does not go ahead.

How do you protect the model and the data?

On phones and laptops the weights are stored encrypted. The key is released only to a device that proves it is genuine and untampered, and where the hardware offers it, the model runs in a sealed-off part of the chip (a TEE). An air-gapped server has no connection to the internet at all. Before it goes live, we red-team the model against attempts to draw it or your data out.

Which training data is needed?

That depends on the task, and it is not always your own data. Sometimes the model learns from your own documents, sometimes from licensed or public datasets, sometimes from carefully generated examples, and often from a mix. Everything is labelled by fixed rules. We hold part of it back as a test set, so we measure quality on data the model has never seen.

What does it cost?

It starts with a feasibility check: together we define the task, the secrecy and where the model must run, and show what a model of your own achieves. Then you get a price proposal based on the scope, with the steps and the computing power the work needs.

Contact

Talk your task through with an engineer.

Tell us briefly which task you want done locally and how strict the secrecy is. One of our engineers thinks it through with you: the model, the training data and where it should run.

You speak directly with a technical specialist.

No account managers or middlemen. Tell us what you are running into and we will show you straight away what works for your situation.

That’s all we need. See also our privacy statement.

Get directions

Search our services, the news and the rest of the site.