MODEL-AS-A-SERVICE

Model-as-a-Service,
delivered as an outcome.

The platform to build, govern and scale AI with confidence. Purpose-built models, trained on your data, running inside your boundary, measured against the outcome you bought.

For regulated industries that need AI to be secure, explainable and sovereign.

The Basinwright console deploying an open-weight model to a realtime endpoint

THE OVERVIEW, CHAPTER BY CHAPTER

What the film walks through.

A 71-second tour of the console, from the problem with rented AI to governance.

01 · The problem

Rented AI

Their models, their prices, your data on their servers.

02 · The five-layer stack

From the chips to the applications

Governed at every layer, in one console.

03 · Compute

One fleet

Your own GPU and CPU servers, pooled as one fleet.

04 · Self-hosted serving

Open models on vLLM

Served on infrastructure you control, to realtime, batch, shadow or edge targets.

05 · Data Hub

Stores, knowledge bases and memory

The data that feeds every model, read where it lives.

06 · Models & agents

Never a black-box proxy

Open-weight and custom models, and the agents built on them.

07 · Projects

AI at work per industry

Dashboards over your own data, one project per team or business line.

08 · Ask Copilot

A live answer from your own workspace

Ask what is deployed, what it costs and how it is doing.

09 · Governance

Access, approvals and audit

Guardrails, an audit trail and per-tenant encryption.

THE PLATFORM UNDERNEATH

You buy the outcome. You keep the platform that delivers it.

The work lands in a console you operate, with your models, your data connections, your deployments and your evidence in it. Not a slide pack and a handover call.

One control plane, every use case

Fraud, claims, procurement and AI safety arrive in the same place, with one audit trail and one access model behind them, rather than six vendors and six governance stories.

Driven by API, not just by console

A REST API across the platform, Python and TypeScript SDKs, an OpenAI-compatible inference gateway that meters every call, and MCP so your models are reachable as tools.

Yours to keep

The estate runs in your tenancy under your keys, and what it produces (models, weights, governed corpora, decision history) stays inside it.

WHAT THE PLATFORM IS MADE OF

Eight capabilities, one audit trail.

Instead of assembling model vendors, GPU providers, a vector store, an evaluation harness and a governance story, you operate one environment, and keep it.

Models as a Service

Production models through secure APIs

Deploy production-ready models without managing infrastructure.

  • Model catalogue
  • Serverless inference
  • Dedicated endpoints
  • Version management
Compute

Enterprise GPU cloud

Capacity for training and serving AI at scale, alongside CPU serving where a GPU is not needed.

  • H100 / H200 / B200
  • Multi-node training
  • Kubernetes
  • Bare metal
Agents

Agents that act under policy

Agents that plan, act and stop for a person when policy says so. 1,141 of them ship ready-made.

  • Memory
  • Tool calling
  • Human approval
  • Knowledge retrieval
Meet the agents →
Studio

Visual AI development

Design, evaluate and ship AI systems without leaving one workspace.

  • Prompt engineering
  • Fine-tuning
  • Evaluation
  • Monitoring
Knowledge

Enterprise RAG

Connect the systems your organisation already runs on and make them answerable, with citations.

  • SharePoint
  • SAP
  • Oracle
  • Salesforce
  • Microsoft 365
Data Hub

Unified enterprise data

Structured, semi-structured and unstructured data in one governed plane.

  • SQL & NoSQL
  • Documents
  • Streams
  • Images & video
Observe

Monitoring and governance

Trace every inference, cost centre and policy decision across the estate.

  • Inference tracing
  • Cost attribution
  • Drift detection
  • Audit export
Marketplace

Models, agents and extensions

Procure vetted AI capability with commercial and compliance terms attached.

  • Vetted publishers
  • Private listings
  • Usage-based terms

MODELS

Deterministic and generative, deliberately both.

Many vendors have reduced every problem to a prompt. That is the wrong tool for most enterprise decisions, and it is why so many pilots die in risk review. Route by cost, latency, residency or capability without rewriting application code.

Model

Deterministic ML

Scorecards, anomaly detectors, survival and time-series models: reproducible and defensible.

Model

Large language models

Reasoning, drafting and synthesis, self-hosted and fine-tuned on your corpus.

Model

Vision

Detection, inspection and document understanding.

Model

Speech

Transcription, diarisation and real-time voice.

Model

Embeddings

Dense and sparse retrieval tuned for your corpus.

Model

Multimodal

Text, image, audio and video in a single context.

Model

Code

Generation, migration and review across your stack.

Model

Domain-specific

Clinical, legal, financial and industrial specialists.

CATALOGUE

A catalogue to start from, not a blank page.

A shipping use case arrives as a working data pipeline, a trained artifact, an evaluation report and a model card. What runs in production is calibrated on your population, your policy and your thresholds.

Fintech & customer

Cut fraud loss without more false declines

Real-time fraud scoring, churn risk, lifetime value and behavioural authentication.

Credit & regulatory risk

Numbers that survive challenge

IFRS 9 ECL, Basel IRB capital, application scorecards and transaction monitoring.

Insurance

Shorter claims cycle time

Automated underwriting, claims triage, fraud-ring detection and loss reserving.

Procurement & supply chain

Recover leaked spend

Spend anomalies, supplier risk, three-way match, tender extraction and spend classification.

AI security & LLM safety

Ship GenAI that passes risk review

Injection, leakage, poisoning and agentic-abuse detection for your own GenAI stack.

  • Shipping: PromptShield
  • LeakSentry
  • PoisonGuard
  • GovernanceLens
Knowledge & GenAI

Answers with sources, not guesses

Grounded assistants and routing over your own document estate, with citations.

  • Shipping: RAGArmor vulnerability scanner
GPU-served & sovereign

Nothing leaves your perimeter

On-premises embeddings, self-hosted LLM assistants and domain classifiers.

Not in the catalogue?

Built to order, scoped to an outcome

Custom models are a productised service on the same pipeline, with the same governance and monitoring.

UNDERNEATH A DEPLOYMENT

We build it, run it and watch it. You own it.

  1. 01 · IngestEverything arrives messy

    Records land unnormalised, duplicated and unlinked, from systems that never agreed on a schema. Nothing has earned meaning yet.

  2. 02 · Cognitive Data HubResolution is where meaning starts

    Entities are resolved and duplicates collapsed into one governed record. What fails the quality bar goes to quarantine, to be fixed or to expire.

  3. 03 · Build & deployWe build the model and deploy it into your estate

    Trained, tuned and evaluated on your governed data, then deployed into your tenancy and your region, under your keys.

  4. 04 · Ground, verify, explainThree engines, running at once

    Retrieval grounds the case in evidence, deterministic models check it against real constraints, and a language model weighs the options and explains the call.

  5. 05 · Run & monitorWatched every hour of every day

    Drift, accuracy, latency and cost per decision are monitored. A model that starts to slip is retrained and re-evaluated before it reaches a decision that matters.

  6. 06 · DecideA case decides only when every lane agrees

    Evidence builds up lane by lane and nothing is decided early. Every decision keeps the lineage of the exact records behind it, and confidence is computed, not asserted.

  7. 07 · OwnYour business owns the model and the intelligence

    The model, the weights, the governed data and every decision it has made stay in your estate. No lock-in.

MODEL 360

An outcome you cannot measure is a claim.

Every model carries the same evidence, on by default: why it decided what it did, a signed trail back to its inputs, and continuous checks that it still holds.

Never a bare number

Contributing factors and reason codes on tabular models, cited source passages on retrieval and language models. Adverse-action reason codes are native, for wherever a declined applicant has a right to an explanation.

Signed reasoning receipts

An Ed25519-signed evidence trail linking each output to its inputs and sources, built for audit and dispute resolution rather than for a dashboard.

Watched, not assumed

Drift and accuracy decay monitored statistically (PSI and CSI), output quality scored by evaluators continuously rather than once at UAT, and every inference logged and metered.

Data-to-decision lineage

Versioned configuration and change history for every model, with validation workflows for the models that carry real risk.

SERVING

Runs wherever your policy allows it to run.

Residency, sovereignty and blast radius are usually decided before the model is. Your business data stays in your systems and is read at query time; the same platform, governance and evidence trail follow the model wherever it lands.

Inside your environment

CPU serving

Scoring, forecasting and classification at a fraction of the cost. Not everything needs a GPU, and we will say so.

Provisioned on demand

Cloud GPU

Embeddings, document intelligence and LLM inference on dedicated capacity.

Inside your perimeter

Your own GPU

On-premises training and inference for sovereignty and residency-constrained work.

Sovereign

Air-gapped

No prompt or embedding data leaves the network. In-country residency where law or contract requires it.

CONNECTED

Connected to the models and systems you already use.

Model providers

Open-weight and frontier, one contract

Route by cost, latency, residency or capability without rewriting application code.

  • OpenAI
  • Anthropic
  • Meta
  • Mistral
  • Google
  • DeepSeek
  • Qwen
  • Microsoft
  • Cohere
  • Open source
Integrations

The systems you already run

Databases, warehouses, object stores, streams and business systems, with credentials sealed at rest.

  • Microsoft 365
  • SAP
  • Oracle
  • Salesforce
  • SharePoint
  • Snowflake
  • Databricks
  • Kafka
  • PostgreSQL
  • ServiceNow
  • Amazon S3
  • Google Workspace
Deployment

Wherever your policy allows

Every option runs the same platform, with the same governance.

  • Cloud
  • Hybrid
  • On-premises
  • Sovereign
  • Private GPU
  • Air-gapped
  • Multi-cloud
  • Edge
Ship with

Your languages and tooling

Anything you can click in the console, you can automate.

  • Python
  • Java
  • .NET
  • JavaScript
  • REST APIs
  • CLI
  • Terraform
  • GitHub

FOR ENGINEERS

Driven by API, not just by console.

Anything you can click, you can automate. Every call routes through the gateway, which is what makes per-team cost and drift monitoring possible in the first place.

Developer documentation →

REST API

A public API across the full platform surface: projects, registry entries, deployments, guardrails and evaluators are all addressable.

SDKs, CLI and Terraform

Python and TypeScript SDKs, a CLI and a Terraform provider, so the estate can be managed as code rather than through a browser.

OpenAI-compatible gateway

Authenticates and meters every call, so usage, cost and behaviour are visible per project, per model and per team.

MCP

Your models and data are reachable as tools by agentic clients, under the same access rules.

MODULES

What you operate day to day.

Every one of these is a surface in the console and an endpoint on the API. Nothing here is on a roadmap.

Projects

Isolated workspaces with their own data, credentials, compute and members.

Data Hub

Register, browse, preview and query the stores your models read from.

Integrations

Databases, warehouses, object stores, streams and AI providers, with credentials sealed at rest.

Models & Registry

Version, configure and promote models, with evaluation gates before promotion.

Deployments

Live endpoints with health, logs, events and rollback.

GPU 360

Fleet inventory, live utilisation and cost per GPU server.

Chat & Knowledge

Grounded answers from your own documents, with the passage attached.

Guardrails & Evaluators

Safety screening on inputs and outputs, and continuous quality scoring.

Automations

Workflows chaining models, data and actions on a schedule or trigger.

Access control

Role- and attribute-based permissions, down to the individual resource.

Billing & usage

Per-project consumption and spend, metered at the gateway.

Agents

Every agent with its state, runs, success rate, cost and outcomes.

PRICING

Start on a laptop. Scale to a sovereign estate.

The platform is priced by what you consume and what you need governed. A delivered use case is scoped separately, against its outcome, because that is work with a target attached, not a seat licence.

Developers · free to begin

Starter

Everything you need to prototype against frontier models.

  • Serverless inference
  • Shared model catalogue
  • Usage analytics
  • Community support
Talk to sales →
Growing teams · most popular

Professional

Dedicated capacity and the controls a real product needs.

  • Dedicated endpoints
  • Fine-tuning
  • Role-based access
  • Evaluation suites
  • Priority support
Book a session →
Large organisations

Enterprise

Private networking, governance and procurement-grade terms.

  • Private cloud & VPC peering
  • SSO, audit and policy
  • Committed GPU capacity
  • Availability terms per estate
  • Named architect
Talk to sales →
Governments

Sovereign AI

National-scale compute with full data and operational sovereignty.

  • Air-gapped deployment
  • In-country residency
  • Cleared personnel where required
  • Bespoke model estate
Talk to sales →

What people ask before procurement does.

How is a delivered use case scoped?

Separately from the platform, and deliberately so. We agree the target metric up front and it becomes the definition of done. We tell you plainly whether the use case ships today as a reference implementation to be tuned on your data, or is built to order. A shipping use case typically goes from kickoff to a governed, monitored endpoint in weeks, and drift monitoring and retuning after go-live are part of the engagement.

Do we pay for GPUs we are not using?

No. Compute is consumed rather than reserved unless you want committed capacity. CPU serving covers a large share of deterministic use cases at a fraction of the cost, and we will tell you when a use case does not need a GPU.

What happens to our models if we leave?

You keep them. The model, the weights, the governed data it learned from and its decisions are inside your estate throughout. Exit terms are agreed up front rather than negotiated under pressure later.

Is the platform priced separately from the delivery work?

Yes. The platform is a subscription with consumption on top; a delivered use case is scoped against its outcome.

Do you hold the certifications our vendor review will ask for?

We describe our controls plainly and do not claim certifications we do not hold. Ask for the current assurance position and you will get the real answer, including where it is incomplete. See Security & governance.

Model-as-a-Service, delivered as an outcome — Basinwright Industries