Try it live Talk to us

Customisation

Built to your job, and to your hardware.

We develop private AI models to the client’s requirement in two dimensions at once: the function it performs, and the machine it has to run on. Because we build the models ourselves rather than fine-tuning someone else’s, we can change the size, context length, vocabulary and capabilities to fit your machine — instead of telling you which machine to buy.

Dimensions: function & hardware

366M

Parameters, tuned per task

You

Own the weights

2–6 wks

Typical build

By function

Trained for the job you actually have.

The model is trained for your task, not adapted from something general. This is work we have built or are building today.

Documents

Document data extraction

Business cards, Emirates IDs, passports, invoices and receipts, tenancy contracts, CVs. Text in, structured JSON out, ready for your CRM or accounting system.

Finance

Accounting head classification

A line item and your own chart of accounts in; the correct head out. Trained on your ledger, not a generic taxonomy.

Data quality

Duplicate detection

Catching the same customer entered three times with three spellings — across scripts, abbreviations and transliterations.

Routing

Sentiment and intent

Routing and prioritising incoming messages against your own category list, not a fixed sentiment scale.

Text

Summarising and rewriting

Meeting notes, long documents and correspondence — condensed, rewritten, tone-adjusted or grammar-checked.

Assistant

Support conversation

A company assistant that knows your products and hands over to a person when it should, answering facts from approved text rather than generating them.

By hardware

Tell us what it has to run on and we design to that constraint.

This is the part few others will take on. The target device is an input to the model design, not an afterthought.

A server you already own

CPU-only is enough for most document workloads. No procurement cycle, no accelerator budget.

A mini PC in the corner

An Intel NUC or N100 box under a desk runs the model comfortably at 4-bit.

A single small GPU

One T4 or L4 class card when throughput or response latency matters more than cost.

A laptop in the field

For teams with no connectivity. The weights are on the disk; the network cable is optional.

An edge or industrial box

A Jetson-class device or ruggedised PC on a factory floor, inside the plant network.

A phone, in time

Under 1 GB of RAM at 4-bit puts in-device inference on modern handsets within reach.

Designed to the device

The hardware is an input to the model, not an afterthought.

Tell us what it has to run on and we change the size, context length, vocabulary and capabilities to fit it. Because we build the models ourselves, we do not have to tell you which machine to buy.

What the model itself needs

366M parameters, so memory depends on precision.

Sizing is not guesswork. Here is exactly what the model occupies at each precision, before any application overhead.

Format	Model size	RAM needed (with context + runtime)
32-bit (what you run today)	1.46 GB	~2 GB
16-bit	0.73 GB	~1.2 GB
8-bit quantised	0.37 GB	~0.8 GB
4-bit quantised	0.18 GB	~0.5 GB

Conversation memory adds only ~100 MB at 1,024 tokens. So the headline number is under 1 GB of RAM, and about 200 MB of storage at 4-bit.

That is the whole argument for small models in one table. A frontier model cannot be quantised into a phone; this one already fits, which is why deployment stops being a negotiation about infrastructure.

What you receive

The weights. The model is yours.

A file on your disk under a one-time licence — no per-user fee, no subscription, no calls to our servers, and no dependency on our company continuing to exist.

The weights themselves. Not API access to a model we host. The artefact is delivered to your infrastructure and stays there.

A one-time licence. Perpetual, per deployment. Cost does not scale with seats, documents or usage.

No outbound call path. You can put it on a machine with the network cable unplugged and it will still work.

A local interface. REST and queue endpoints inside your network, returning structured JSON your systems can consume directly.

Re-training rights. Fine-tuning on new formats and edge cases is part of the licence, not a new purchase.

By function

The work we build

Documents

Extraction

Emirates IDs, passports, visiting cards, invoices, receipts, tenancy contracts and CVs. Text and images in, named fields out against a schema you define — with a confidence value and an explicit list of anything the model could not read.

Finance

Classification

A line item and your own chart of accounts in, the correct head out — trained on your ledger rather than a generic taxonomy, so it matches how your finance team actually codes. Duplicate detection runs on the same principle across spellings and scripts.

Assistant

Support conversation

A company assistant that answers from your approved product text rather than generating claims, routes by intent against your own category list, and hands over to a person the moment it is out of scope.

How an engagement runs

Scoped against your real documents, not a demo set.

STEP 1

Scope the task

We define one workflow, the formats involved and what “correct” means for it — measurable, not impressionistic.

STEP 2

Review real data

We work from your actual archive, with personal details substituted so layouts are learned and people are not.

STEP 3

Train and evaluate

The model is built and measured against held-out real documents. Failures are reported as carefully as successes.

STEP 4

Deploy and hand over

Installed on your hardware, wired into your systems, with the weights, the licence and the runbook in your hands.

What we will not claim

A narrow tool, honestly described.

Not this

A general assistant

REX42 does not compete with frontier systems on open-ended reasoning, world knowledge or long-form writing. If that is the requirement, a large hosted model is the right tool and we will say so.

This

A dependable specialist

For a defined task, on your own hardware, with your data never leaving the building, a small well-trained model is the better tool — and often the only permissible one.

Keep reading

Deployment Every device it runs on, and what it costs to host. Data Security Why local inference removes the question entirely. Playground Test the current capabilities on your own files.

Tell us the task and the hardware.

We will size it honestly against your real documents before you buy anything.

Talk to us See deployment options