AI systems
AI that does real work — inside your systems.
We design, build and support AI systems that are scoped to an outcome, grounded in your data, wired into the tools your people already use, and governed from day one. Six capabilities, one engineering team.
Capability explorer
Pick a capability. See how it works.
Most production systems combine several of these. An assistant might use retrieval for answers, an agent for actions and automation for the paperwork behind both.
Custom AI systems
An end-to-end system built around one process or decision: data in, model reasoning in the middle, a useful action or interface out — with evaluation and monitoring wrapped around it.
Typical fit: triage, forecasting, recommendation or review processes where staff spend hours on repeatable judgement calls.
Custom AI systemsLLM applications
Assistants, copilots and natural-language interfaces on large language models — with prompt versioning, structured outputs, evaluation suites and cost controls so they behave consistently in production.
Typical fit: internal assistants, drafting tools, customer-facing chat, natural-language search over structured data.
LLM applicationsAI agents
Agents that plan, call tools and APIs, and complete multi-step tasks. Every tool has a defined permission scope; consequential actions wait for human approval; every step is logged.
Typical fit: back-office workflows that span several systems — reconciliation, follow-ups, case preparation.
AI agentsRAG & knowledge systems
Retrieval-augmented generation over your policies, documents and records. Answers cite their sources, and retrieval respects the same permissions as the systems the content came from.
Typical fit: policy and procedure Q&A, technical documentation, case history, tender and contract libraries.
RAG & knowledge systemsModel integration & fine-tuning
Choosing the right model for the job — hosted, private or open-weight — integrating it cleanly, and fine-tuning only where measured results show it beats retrieval and prompting.
Typical fit: data-residency constraints, specialised language or formats, high-volume tasks where cost and latency matter.
Model integrationAI automation
Document processing, classification, extraction and routing that removes manual re-keying between systems — with confidence thresholds that send uncertain cases to a person.
Typical fit: invoices, applications, forms, inbound email and anything else that arrives as unstructured text.
AI automationReference architecture
Five layers. One governed system.
Whatever the capability, the system has the same shape, which keeps it testable, replaceable and auditable. Models change quickly; the layers around them shouldn’t have to.
- Swappable models. The model layer sits behind an interface, so moving providers is a configuration change plus re-evaluation.
- Grounded outputs. Retrieval and tools connect answers to your data instead of the model’s memory.
- Governance everywhere. Evaluation, guardrails, access control and audit logging apply to every layer.
- L1InterfacesWeb apps · mobile · chat · APIs
- L2Agents & orchestrationPlanning, tools, workflows, approval steps
- L3Knowledge & retrievalRAG over documents, records and data
- L4ModelsHosted, private or fine-tuned LLMs
- L5Your systems & dataCRM, ERP, databases, files
What “production” means here
The parts a demo leaves out.
Evaluation harness
A test set built with your team, run on every change to prompts, models or retrieval.
Observability
Traces of each request — inputs, retrieved sources, tool calls, outputs, latency and cost.
Access & secrets
Least-privilege service accounts, scoped tool permissions and managed secrets.
Run & improve
Monitoring for drift, failure modes and spend, with a backlog of improvements after launch.
FAQ
Common questions.
Straight answers to what teams usually ask before starting an AI project.
Which AI models do you use?
We’re model-agnostic. We evaluate hosted models from the major providers and open-weight models you can run privately, then choose on measured quality, cost, latency and data-residency requirements for your use case.
Does our data get used to train someone else’s model?
We design so that it doesn’t: we use provider settings and agreements that exclude your data from training, or private deployments where the data never leaves an environment you control.
How do you stop an AI system making things up?
By grounding answers in retrieved sources with citations, constraining outputs to defined formats, testing against an agreed evaluation set before release, and routing low-confidence cases to a person.
Can you work with our existing systems?
Yes. Most of our work is integration: connecting AI to CRMs, ERPs, document stores, databases and line-of-business applications through their APIs, with the least access needed.
How do projects usually start?
With a consultation to understand the outcome you want, then a short scoping phase to confirm data, risks and success measures before committing to a build.
Have a process that should be smarter?
Tell us about the workflow, the data and the outcome. We’ll suggest a practical first step.