swfte
Kontakt

Platform / Custom models

Build models for your domain

Models adapted on your own data, evaluated before they ship and deployed under your control, with safety first and EU hosting in mind.

A general model knows a little about everything and nothing about how your organisation works. A custom model is adapted to your domain, your vocabulary and your task, and it runs where you decide. This page sets out when that is worth doing and when retrieval or a better prompt is the better answer. It then walks the seven stages from picking a base to retraining, and it says plainly which stages exist today. Hosting and serving your own weights is built. Managed adaptation, and selecting training data from the company brain, are on the roadmap.

Build with AITalk to our team

What a custom model is

A custom model is an open-weight model that has been adapted to your domain. It might be a small model taught the shape of your tickets, a drafting model that writes in your house style, or a classifier that knows your alert vocabulary. The adaptation can be a light adapter such as LoRA, or a fuller fine-tune. The result is a set of weights that you can host, version, test and retire like any other asset.

It is a way of changing behaviour, not a way of storing facts. A fine-tune is good at tone, format, vocabulary and reliability on a narrow task. It is poor at holding facts that change, which is the job of retrieval and the company brain. Most good systems use both: the model supplies the behaviour and the brain supplies the facts.

Swfte treats this as a pipeline with gates, not a one-off training run: pick a base, prepare data, adapt, evaluate, harden, deploy, monitor and retrain. Each stage is described below with its honest status, and the same safety-first method applies to every model, including our own.

Prompt, retrieve, fine-tune or train: when each earns its cost

The cheapest lever that works is the right one. Move down the table only when the one above has been tried and measured.

When to prompt, retrieve, fine-tune or train a model, and what each needs
LeverReach for it whenWhat it needsMain riskSwfte today
PromptingThe task is general and a clear instruction and a few examples get the behaviour you want.A good prompt, a few worked examples and an evaluation set to catch regressions.Fragile when the model or the prompt changes. Long prompts cost on every call.Built. Studio and Connect.
Retrieval, or the company brainThe answer depends on facts that change or that belong to you: policies, owners, documents, history.Sources connected, access rules applied and a way to check the answer against its source.Retrieving the wrong passage, or one the asker should not see.Directory graph built. Document search in progress. Other sources on the roadmap.
Fine-tuning or adapting a modelYou need a stable behaviour, format, tone or narrow-task reliability that prompts and retrieval do not give you, often at lower cost per call.Clean examples of the behaviour, a held-out evaluation set, an owner, and a safety re-test after every change.Overfitting, forgetting, and eroded safety behaviour. Facts baked into weights go stale.Hosting and serving your weights is built. Managed adaptation is on the roadmap.
Training from scratchThe deliverable is a licence and a data lineage you fully control, not a behaviour.A large, licensed corpus, serious compute and a team that has done it before.Cost and time, with a weaker result than a good open-weight base for most tasks.Research only. Not offered.

The rule of thumb is short. Fine-tune for behaviour. Retrieve for facts. Train from scratch only when the thing you need is a licence and a data lineage rather than a behaviour. If you cannot yet say how you would measure the improvement, you are not ready to fine-tune.

Domain fine-tuning in depth

The pipeline: seven stages, each with a gate

The stages run in order, and the last one feeds the first. A failure at any gate stops the run. Each stage carries its status, because the honest answer is that some exist and some do not.

The same seven stages are listed in order below, with their status.

Legend

  • Built. Exists today and can be used.
  • In progress. Being built. Not yet something to rely on.
  • Roadmap. Designed for and on the roadmap. Not built. No dates are given.
  1. 01

    Pick a base

    Built

    Choose an open-weight model by task, licence, size, language coverage and the hardware you will serve it on. Record the exact revision.

    Today: Swfte publishes the method for choosing and testing open-weight models. <supported base models - founder to fill>

    Read more about Pick a base
  2. 02

    Prepare data from the brain

    Roadmap

    Select examples from the company brain by purpose and by access, sanitise them, record consent and purpose, and split them so evaluation data never leaks into training.

    Today: The sanitisation gateway is in progress. Selecting training data from the brain is designed for and not built. You can prepare data with your own tooling today.

    Read more about Prepare data from the brain
  3. 03

    Adapt

    Roadmap

    Adapt the base to your domain with a LoRA adapter or a fuller fine-tune, with the run configuration and data lineage recorded.

    Today: Managed adaptation runs are designed for and not built. You can train with your own tooling and bring the weights. <supported adaptation methods - founder to fill>

    Read more about Adapt
  4. 04

    Evaluate

    In progress

    Score the candidate on held-out domain tasks and on safety, and compare it with the model in production. A failing gate blocks promotion.

    Today: Single-turn evaluation exists in Studio and the method is published. Named domain suites are designed for. <named evaluation suites - founder to fill>

    Read more about Evaluate
  5. 05

    Harden

    In progress

    Red-team the model, check the weights against pinned hashes, set runtime guardrails and fix the policy it runs under.

    Today: Content-policy checks with redaction run at the gateway. Red-teaming follows the published method. An automated red-team pipeline is designed for.

    Read more about Harden
  6. 06

    Deploy

    Built

    Serve the model on dedicated infrastructure behind the Connect gateway, promote it through stages, and keep the previous approved revision routable.

    Today: Upload weights, version them, promote dev to staging to production, deploy to a dedicated endpoint and route through Connect, with an audit log. On-premises and air-gapped serving are designed for or on request.

    Read more about Deploy
  7. 07

    Monitor and retrain

    In progress

    Watch quality, safety signals, latency and cost on live traffic, write outcomes back to the brain, and retrain when the evidence says the model has drifted.

    Today: Usage, cost and latency are visible through Connect. Drift detection, outcome write-back and retraining triggers are designed for.

    Read more about Monitor and retrain

Safety first, in the method and in our own model

Fine-tuning can make a model worse at refusing things it should refuse, even when the training data is harmless. So every adaptation is treated as a new candidate and is tested again for refusal, over-refusal, prompt-injection resilience and behaviour in EU languages, against the model it would replace. A model that fails a gate does not ship.

Swfte Safety is the working name for Swfte’s own safety-first fine-tuned model, designed for governed agents and regulated workflows. Its design intent and model card are published. No measured results are published yet, and we do not publish numbers we cannot reproduce. The same method that applies to open-weight models applies to it, after every change.

Fine-tuning for safety lowers the rate of harmful behaviour. It does not remove it. Runtime guardrails, access control and human approval remain part of the design around any model, custom or not.

The Swfte Safety model card · How we test open-source models · Evaluation and safety in depth

EU first: where it runs and what it starts from

Models run on dedicated, single-tenant infrastructure in a region you choose, with options from an isolated VPC to hardware in your own data centre. The exact regions and facilities are agreed per engagement. <EU regions and facilities offered - founder to fill>

Starting from an open-weight base lets you keep both the weights and the data inside your boundary. The base still has a licence. Licences differ per checkpoint and change between releases, and some restrict use in the EU, so read the licence that ships with the exact weights and prefer permissive terms such as Apache 2.0 or MIT where the model quality allows.

Swfte provides the technical controls, governance mechanisms and evidence required to deploy AI within an organisation's applicable regulatory, security and policy requirements. The exact posture depends on the customer's use case, jurisdiction, deployment and configuration. Swfte does not hold a SOC 2 report, an ISO 27001 certificate or a HIPAA BAA today. A SOC 2 Type I audit is in preparation; the trust page lists what is in place and what is in progress.

Deploy models on infrastructure you control · Deploy and serve in depth

Who owns what

Ownership is the point of building your own model. These are the six things to be clear about before any work starts.

  • Your weights

    The adapter or weights that come out of the work are designed to be yours to keep, move and run elsewhere, subject to the licence of the base model they were built on and to your agreement.

  • Your data

    Training data stays inside your boundary by design. It is selected from your own brain and sanitised first, and the exit and export terms are set in your agreement.

  • Your evaluation sets

    The held-out cases that decide whether a model ships belong to you, and are kept out of training so a good score means something.

  • Your lineage

    Base model and revision, data sources, evaluation results and approvals are recorded together, so you can say what ran and why.

  • Your switch

    You decide who can promote a model and who can retire it. Retiring a model, or erasing the data behind it, is your decision and the record shows it.

  • What is not yours

    The base model. Its licence still governs what you may do with derivatives, so read it for the exact checkpoint, and have counsel read it before an EU deployment.

Own your model in depth

What exists today, and what is designed for

Built means it exists and can be used. In progress means it is being built. Roadmap means it is designed for and not built. We give no dates.

What is built, in progress and on the roadmap for custom models
CapabilityStatusNotes
Host your own model weights: upload, hash manifest, versions, promotion through dev, staging and production, audit logBuiltModel Vault in Studio. Plan availability: <custom model availability and plans - founder to fill>
Deploy a model to a dedicated endpoint and route to it through the Connect gatewayBuiltConnect puts one OpenAI-compatible API in front of the endpoint, with routing, failover, usage caps, cost tracking and audit events.
Content-policy checks with secret and personal-data detection and redaction at the gatewayBuiltApplies to requests through Connect, including those to your own model.
Single-turn evaluation of model outputBuiltScores whether a model said the right words. It does not score what an agent with tools does.
A published method for testing open-weight modelsBuiltCapability, safety, multilingual, licence and supply-chain checks with regression gates. Results are published only when reproducible.
Swfte Safety, a safety-first fine-tuned modelIn progressDesign intent and model card are published. No measured results are published yet.
Pre-model sanitisation gatewayIn progressDesigned to remove what must not reach a model, including from training data.
Agent evaluation sandbox with recorded, replayable tracesIn progressInternal tooling today. Not offered as a customer feature.
Selecting training data from the company brain, with purpose and consent recordedRoadmapDesigned for. The brain does not export training sets today.
Managed adaptation runs: LoRA or fuller fine-tuning on infrastructure you controlRoadmapDesigned for. A small proof-of-concept pipeline has been run internally. It is not a service. <regions where training runs - founder to fill>
Named domain evaluation suites and automated red-teamingRoadmapDesigned for. <named evaluation suites - founder to fill>
Drift detection, outcome write-back to the brain and retraining triggersRoadmapDesigned for.
Serving in your own cloud, on-premises or air-gappedRoadmapDesigned for, and on request. See Connect self-deploy for what exists.

Legend

  • Built. Exists today and can be used.
  • In progress. Being built. Not yet something to rely on.
  • Roadmap. Designed for and on the roadmap. Not built. No dates are given.

Two examples, described the way every governed agent is

Each shows what an agent on a custom model can do, what it cannot, what needs approval and what it records. Each also says which parts depend on what is built.

A drafting assistant on your own domain model

In progress

“Draft a response in our house style to this supplier query, citing the relevant clause.”

A small model adapted to your house style and vocabulary drafts. Retrieval supplies the clause. A person approves before anything leaves.

Regulated Drafting Assistant

Can
  • Draft in the organisation’s tone and format
  • Quote clauses retrieved from documents the asker may open
  • Flag when it is unsure and ask for a person
Cannot
  • Send anything externally
  • Decide what a clause means or give legal advice
  • Use facts from its training data in place of the cited source
Requires approval
  • Sending the draft
  • Any commitment to a supplier
  • Any change to a contract record
Records
  • Who asked
  • Model and revision used
  • Sources cited
  • Draft produced
  • Approver
  • Outcome

Built today: Hosting and serving your own weights through the vault and Connect, with policy checks and an audit log.

Designed for: Selecting the style examples from the brain, managed adaptation and the domain evaluation gate.

Alert triage on a model that knows your estate

Roadmap

“Is this alert one we have seen before, and who handles it?”

A model adapted to the organisation’s alert vocabulary classifies and summarises. The brain supplies the owner and the on-call approver. Containment waits for a person.

Alert Triage Assistant

Can
  • Classify and summarise alerts in the organisation’s own terms
  • Name the owner and approver the brain resolves, with evidence statuses
  • Open an incident with the evidence attached
Cannot
  • Disable a service, revoke a credential or change a policy
  • Treat a stale or disputed owner as settled
  • Read outside the incident scope
Requires approval
  • Containment of any kind
  • Notifying anyone outside the owning group
  • Closing an incident as benign
Records
  • Alert
  • Model and revision used
  • Owner and approver resolved
  • Summary
  • Decision
  • Approval
  • Outcome

Built today: Serving a model behind Connect, and Trust Profiles and approval rules for the agent that uses it.

Designed for: Service ownership from the brain, the data selection and adaptation steps, and the action gateway.

The loop: brain, model, agents, outcomes, back into the brain

A custom model is not a destination. It is one station in a loop. The brain supplies the evidence the model is adapted on, agents use the model under policy, their outcomes return to the brain as new evidence, and that evidence decides when the model needs another round.

The same four stations and four arrows are listed in order below.
  1. 01Company brainHolds what the organisation knows, with evidence statuses, history and access rules.
  2. 02Custom modelAdapted on data chosen from the brain, then evaluated and hardened before it ships.(this page)
  3. 03Governed agentsUse the model and read the brain, inside a Trust Profile, with approval where it matters.
  4. 04OutcomesWhat happened: approvals, corrections, results and cost, all on the record.

The four arrows

  1. Company brain to Custom model: select, sanitise, adaptRoadmap

    Choose training data from the brain, remove what must not reach a model, adapt an open-weight base. The sanitisation gateway is in progress, and the data selection and training steps are on the roadmap.

  2. Custom model to Governed agents: serve, governBuilt

    Serve the model on dedicated infrastructure behind the Connect gateway and bring agents onto it under policy. Model hosting and the gateway are built.

  3. Governed agents to Outcomes: act, recordBuilt

    Agents act within their Trust Profile, with human approval for consequential steps, and every action is recorded.

  4. Outcomes to Company brain: written back as evidenceRoadmap

    Outcomes return to the brain as new evidence with a status, and they decide when the model needs retraining. The write-back is on the roadmap.

Legend

  • Built. Exists today and can be used.
  • In progress. Being built. Not yet something to rely on.
  • Roadmap. Designed for and on the roadmap. Not built. No dates are given.

The company brain

Go deeper

Five pages, each with its own question.

  • Domain fine-tuning

    A plain guide to choosing between prompting, retrieval, fine-tuning and training from scratch, and to adapting an open-weight base on your own data without losing safety along the way.

  • Data preparation

    How to choose training data by purpose and by access, remove what must never reach a model, keep evaluation honest, and keep a record you can defend later.

  • Evaluation and safety

    How a custom model is tested on real domain tasks and on safety before it serves anyone, and why every adaptation is tested again from the start.

  • Deploy and serve

    How a custom model gets from a weights file to a governed endpoint: the Model Vault, the Connect gateway, promotion, rollback and the policy around every request.

  • Own your model

    What owning a custom model actually means, what you do not own, and how portability, control, sovereignty and retirement work in practice.

Specifics we have not published yet

We would rather leave a gap than invent a detail. These are for the founder to fill before they are stated on the site.

Supported base models
<supported base models - founder to fill>
Adaptation methods
<supported adaptation methods - founder to fill>
Regions where training runs
<regions where training runs - founder to fill>
Named evaluation suites
<named evaluation suites - founder to fill>
Availability and plans
<custom model availability and plans - founder to fill>
Engagement model
<self-serve or engagement model - founder to fill>
Pricing
<pricing - founder to fill>

Frequently asked questions

When should I fine-tune a model instead of using RAG?

Fine-tune when you need a stable behaviour, format, tone or narrow-task reliability that prompts and retrieval do not give you. Use retrieval for facts that change or that belong to you. Many systems use both. If you cannot say how you will measure the improvement, start with a better prompt and retrieval.

Does Swfte fine-tune models for me today?

Not as a managed service. Hosting and serving your own weights is built: upload, version, promote and deploy through the Model Vault, and route through Connect. Managed adaptation runs are on the roadmap. <self-serve or engagement model - founder to fill>

Which base models can I start from?

Open-weight models whose licence your review clears and that fit your hardware. We publish how we test them. <supported base models - founder to fill>

Can I train on data from the company brain?

That is the design: data selected from the brain by purpose and by access, sanitised first, with consent recorded. It is on the roadmap, and the sanitisation gateway is in progress. Today you can prepare data with your own tooling and bring the resulting weights.

Do I own the resulting model?

You own your weights or adapter, your data and your evaluation sets, subject to the licence of the base model underneath. The base model licence still governs derivatives. Read it for the exact checkpoint.

Does fine-tuning make a model less safe?

It can, even with harmless data. That is why every adaptation is evaluated again for refusal, over-refusal and prompt-injection resilience, and compared with the model in production before it can ship.

Where does the model run?

On dedicated infrastructure in a region you choose, behind the Connect gateway. On-premises and air-gapped serving are designed for or on request. <EU regions and facilities offered - founder to fill>

What is Swfte Safety?

Swfte Safety is the working name for Swfte’s own safety-first fine-tuned model. Its design intent and model card are published. Its base model, data provenance, results, release date, licence and availability are still to be published, and no measured results exist yet.

Is a custom model compliant?

No model is compliant on its own, and we do not claim it. Swfte is built for compliance-by-design: it provides technical controls, evaluation evidence and an audit trail that help you meet your own obligations. The exact posture depends on your use case, jurisdiction, deployment and configuration. Swfte does not hold a SOC 2 report, an ISO 27001 certificate or a HIPAA BAA today. A SOC 2 Type I audit is in preparation; the trust page lists what is in place and what is in progress.

What does it cost?

<pricing - founder to fill> <typical project timeline - founder to fill> Talk to our team for the current position.

Across every layer

These ideas apply to every layer of the platform.

Build a model for your domain, and keep it

Start with one entry point. Add intelligence, agents, workflows and infrastructure as you prove value.

Build with AITalk to our team

Deploy a model with Swfte Connect

One gateway, every provider, per-token cost visibility. Swap models without touching your code.

Get an API keyLearn more