Navitra
NewNaviKinase 1.0 evaluation report

Sovereign AI.

Pre-trainedPrivateOwnedSupported

Explore our models Book a call
Our models

Meet the Navitra models.

Every model is pre-trained from scratch, with no borrowed base model, and runs where its users need it.

Compare all models
Financial crime
Available

NaviDetect 1.0

A language model for anti-money-laundering and sanctions work.

Built to support a financial crime analyst’s judgement, not replace it. Every output is logged so a case can be reconstructed for audit.

Type
Language model, pre-trained from scratch
Built for
AML and sanctions analysts
Runs in
Your infrastructure; weights owned by you
Learn moreRequest access
Cybersecurity
Available

NaviProtect 1.0

A mixture-of-experts model for security teams.

Different expert networks specialise in different parts of security work. Agents built on it ask a person before any irreversible action.

Type
Mixture-of-experts language model
Built for
Security operations teams
Runs in
Your own GPUs or private cloud
Learn moreRequest access
Mortgage underwriting
Available for evaluation

Mortgage Agent

The agent prepares the case. The underwriter decides.

Reads the application file, checks it against your lending policy and returns a case summary with every figure traced to its source page.

Type
AI agent on a Navitra language model
Built for
Mortgage underwriters and lenders
Runs in
The lender’s infrastructure
Learn moreRequest the brief
Protein science
Published

NaviKinase 1.0

A protein language model for the kinase family.

149.6M parameters, pre-trained from scratch on 2.45 million sequences. Matches ESM-2 and ProtGPT2 at identifying enzyme class without labelled examples.

Type
Protein language model, 149.6M parameters
Built for
Protein scientists and researchers
Weights
Public on Hugging Face, CC BY-NC 4.0
Learn moreRead the paper (PDF)
Who uses them

Built for the people who do the work.

Each model is narrow on purpose: trained for one job, for the people who do that job, and evaluated against what they need before it ships.

Underwriters

Spend your time on judgement, not paperwork.

The Mortgage Agent assembles and checks the file. You see the evidence behind every figure and make the call.

Mortgage Agent
Financial-crime analysts

Review each case with the evidence in front of you.

NaviDetect 1.0 supports AML and sanctions reviews, and every output is logged so a decision can be reconstructed.

NaviDetect 1.0
Security teams

Keep incident data inside your walls.

NaviProtect 1.0 runs on your own GPUs or private cloud, so logs never go to a third-party API.

NaviProtect 1.0
Scientists

Score and generate kinase sequences on your own machine.

NaviKinase 1.0 runs locally on Apple Silicon, with its weights and full evaluation published.

NaviKinase 1.0
For organisations

Your own model, trained on your data.

We pre-train a model on your corpus, build the agents around it and run it inside your infrastructure. The weights are yours.

Book a callHow we work
For researchers

Use NaviKinase 1.0 today.

The weights are on Hugging Face under CC BY-NC 4.0, with an evaluation report setting out what the model can and cannot do.

Get the weightsRead the paper
What sovereign means here

We pre-train language models from scratch, on your data, for your domain. Not a wrapper around someone else’s API, and not a borrowed open-source base with a new label. The architecture, the training run and the weights are yours, running inside your own infrastructure. We build the agents on top, then stay on to evaluate, retrain and support them.

Model
Trained for you, from scratch
Weights
Yours to keep
Runs in
Your cloud, on-prem or London
Data protection
UK GDPR · ICO registered
After go-live
Monitoring, retraining, support
Shared with other customers
Never
Latest

News and research

All research

A model is not a deliverable. It drifts, your business moves, and the system that impressed the steering committee in March is quietly wrong by September.

Most suppliers hand over weights and leave. We stay on the evaluation set, the retraining cadence and the support line, because that is where the value either holds or evaporates.

How we build them

Capabilities / 01—06
01

Trained from scratch

We pre-train on your corpus rather than calling someone else’s API or dressing up a general model. Sized to the job rather than to a leaderboard — from 150M for a narrow sequence model up to 15B for broader work.

02

Owned outright

The weights, checkpoints and evaluation sets are yours under the contract. No licence inherited from someone else’s base model, and no vendor who can change the model underneath you.

03

Agents that act on it

We build the agents, tools and approval gates around the model so it acts inside your systems, with an audit trail compliance can read.

04

Evaluated before it ships

We agree the evaluation set before training, measure against it on held-out data, and release the checkpoint that tests best — not the one that finished last.

05

Maintained and retrained

Monitoring in production, a retraining cadence agreed up front, and a support line. A model without this is a depreciating asset from the day it lands.

06

Advisory before you commit

Feasibility, build-versus-buy, model sizing, data readiness and evaluation design. Sometimes the honest answer is that you do not need a bespoke model, and we would rather say so early.

Built to pass your security review

Our governance and AI policies
Deployment

Your VPC or on-prem

Your infrastructure team keeps the keys and the network boundary.

Control

Human-in-the-loop and audit trails

Agents pause at approval checkpoints, and every action is logged for compliance to read.

Data protection

Registered with the ICO

Registration ZC079787. UK GDPR controls are set out in our privacy notice. We will complete your security questionnaire or sign your DPA.

Data

Your data trains only your model

Nothing from your corpus reaches another customer’s model, and the weights are yours to keep.

Things you can check

See it on your own data

Book a free 30-minute call or email contact@navitratech.com · reply within 24 hours · UK based