Proprietary models

In development

Compact text and speech models for specific business tasks. Built towards a practical goal: running useful AI within an environment your organisation approves.

We’re developing proprietary models with a focus on task quality, computing requirements and control over data processing. Once developed and validated, these models are intended for self-hosted deployment, operated by your team or managed by SAMVIT Labs.

Explore the model directions
An illustration of the intended deployment. The environment and operating arrangements are agreed for each project.

Model development directions

These five directions define the work we’re pursuing. Examples below illustrate intended behaviour; they are not outputs from live models or reported performance results.

Text models

Compact LLMs and SLMs, scoped around a task, its language and the information it needs.

Extraction and classification

Structured text

Small language models that turn business text into defined fields. A useful starting point for requests, records and documents with a clear structure.

Explore custom AI agents
Development direction
What we tailor
Your terminology, field definitions and representative examples.
What we evaluate
Field accuracy, valid output and recognition of missing information.
Illustrative example

A service request

Order 4821 arrived with two damaged units.
Order reference
4821
Issue
Damaged items
Units affected
2

Unstated fields, such as delivery date, should remain unknown.

Context-based responses

Domain language

Focused language models for drafting answers from supplied business knowledge. Training can shape terminology and response behaviour; retrieved material supplies current facts.

Explore chatbots
Development direction
What we tailor
Domain vocabulary, response style, source use and escalation rules.
What we evaluate
Factual support, useful answers and appropriate requests for review.
Illustrative example

A support question

Can I return an opened item?
Reference material
The supplied returns policy
Target response
An answer supported by the relevant policy passage
When the policy is silent
Ask for clarification or refer for review

Answers should cite the supplied policy and request review when it does not resolve the question.

Language-specific coding assistance

Focused code

Code models scoped to a programming language and a defined development task. One starting point is Python test generation for a chosen framework and repository.

Explore agent workflows
Development direction
What we tailor
Language, libraries, repository conventions and a bounded task.
What we evaluate
Tests that run, detect faults and follow the documented behaviour.
Illustrative example

Python / pytest

Draft a test for a function that trims spaces and uppercases an order reference.
Target assertion
normalise_reference(" ab-42 ") == "AB-42"
Next step
Run the test against the implementation
Review
A developer checks the test and its assumptions

The model proposes code. A coding agent adds tools, execution and workflow controls around it.

Voice models

Specialised speech models, starting with a selected language, a clear purpose and representative audio.

Language-specific speech to text

Speech recognition

Speech models focused on a selected language and business setting. The work starts with representative speakers, relevant terminology and the conditions in which audio is recorded.

Explore speech to text
Development direction
What we tailor
A chosen language, domain vocabulary and recording conditions.
What we evaluate
Transcription errors and accuracy on names, references, dates and amounts.
Illustrative example

A service-call transcription task

A caller gives an order reference, explains an issue and asks for a follow-up.
Target output
A readable transcript in the selected language
Critical details
Preserve order references, names and numbers
Unclear audio
Flag uncertainty for review

The selected language, vocabulary and recording conditions define the evaluation set.

Language-specific text to speech

Speech synthesis

Speech models for clear, purposeful audio in a selected language. Initial tasks can be short instructions, service confirmations and spoken application responses.

Explore text to speech
Development direction
What we tailor
Pronunciation, speaking style and appropriately licensed voice data.
What we evaluate
Intelligibility, names and numbers, listener feedback and generation time.
Illustrative example

A spoken confirmation

Your appointment is on 14 March at 10:30. Your reference is AB-42.
Target output
Spoken audio in the selected language
Details to preserve
The date, time and reference
Listener review
Check pronunciation, pacing and intelligibility

Pronunciation and pacing need listener evaluation before audio samples can represent a released model.

Deployment and data control

Self-hosting gives an organisation more choice over where model processing happens. Who operates the system and who can access data are separate decisions. We’re designing towards these deployment options as the models mature.

01

Client operated

Your team runs the model within your approved infrastructure.

Data & access
Your team administers the environment and controls access.
Operating cost
Allow for hardware or cloud capacity, maintenance and internal expertise.
02

SAMVIT Labs managed, in your environment

Your organisation controls the hosting account. SAMVIT Labs operates the system with agreed permissions.

Data & access
Hosting ownership stays with you; operational access is explicitly defined.
Operating cost
Infrastructure costs sit alongside an agreed management service.
03

SAMVIT Labs hosted

SAMVIT Labs runs the model in an environment agreed with your organisation.

Data & access
SAMVIT Labs processes data in this option. Location, isolation, access and retention need clear agreements.
Operating cost
Managed operation reduces your internal workload. Dedicated capacity can cost more at low usage.

Data location, administrative access, logging and retention need to be defined for the whole deployment. Self-hosting is one part of that design. Supported configurations will be documented with each evaluated model.

Efficiency has to
hold up in use.

A smaller model can need less computing capacity. The actual cost also depends on traffic, simultaneous users, response times and how much hardware sits idle. Our goal is to make private deployment practical for the workload it serves.

Methods we evaluate

Task-specific training
Shape a suitable starting model around a defined task and representative examples.
Distillation & quantisation
Explore smaller students or lower-precision weights, then check the effect on task quality.
Serving around the workload
Evaluate batching, reuse and capacity sizing where they fit. Live speech and overnight document processing have different needs.

Evaluation criteria

Task quality
Accuracy and failure cases on a defined test set.
Memory & hardware
Tested equipment, model size and memory required to run it.
Responsiveness
Response time and throughput at a stated number of simultaneous requests.
Operating cost
Cost per completed task or minute of speech, including hosting and operation at a stated workload.

These are evaluation criteria. Model-specific measurements will be shared as the work is validated.

Evaluation and releases

Useful development starts with a baseline and a way to recognise an improvement. Quality, deployment requirements and cost need to be assessed together before a model enters a client workflow.

  1. Define the task

    Agree the output, representative examples, sensitive fields and the conditions in which the model will be used.

  2. Establish a baseline

    Evaluate a suitable starting model and record quality, response time and resource use on the same workload.

  3. Adapt and evaluate

    Use task-specific training or compression where appropriate. Test on held-out examples and inspect failures.

  4. Document the release

    Record intended use, tested hardware, results, limitations, licensing and access options before release.

Programme in development

Proprietary model releases are still ahead.

Model downloads and release packages are not available on this page yet. Evaluated releases will describe intended use, training approach, tested hardware, results and limitations, together with licensing and access options.

Our custom client solutions use established models.

A few practical questions.

Can you build a solution while these models are in development?

Yes. Our custom client solutions use established models today. The proprietary model programme is intended to extend the deployment choices available for future projects.

Will a smaller model run on our existing hardware?

That depends on the model, workload and response-time requirements. Hardware suitability needs to be tested for the intended use. Each evaluated model will include its tested requirements so a deployment can be assessed against your environment.

Who would have access to our data?

That follows the operating arrangement. A client-operated deployment and a SAMVIT Labs-hosted deployment have different access requirements. Hosting location, operational permissions, logging and retention need to be agreed before deployment.

From the journal.

Browse the journal

Research and engineering context for this work.

Have a task
or language in mind?

Tell us what the model needs to do, the examples you can provide and where it would need to run.

Discuss your use case