Skip to content
cirlume.
Menu
← All expertise

Expertise / 03 of 04 / Applied AI

AI that does real work, built in the UK.

Cirlume is a UK software studio based in Holywell, North Wales. We design and build AI features that take on specific jobs inside your business: reading documents, answering staff questions from your own policies, routing requests and drafting replies. A person stays in charge of the decisions that matter.

Describe the task
A laptop showing a scanned document beside the fields extracted from it, some highlighted in coral, next to a stack of invoices

We test every AI feature on real examples from your work before it goes live, show the sources behind each answer, and log what the model did. When ordinary software or a person would do the job better, we say so.

We work with businesses across the UK and the Middle East, in health and social care, medical education, property and fintech. We also run products of our own, SoulCare and SimRhythm, so we know what software needs after launch.

When to bring us in

Sound familiar?

  • Your team spends hours searching documents or copying data between systems.
  • You want to try AI on a real task without betting the product on it.
  • You need AI output that someone can check before it is acted on.

Why us

What we bring to applied ai.

  • A person stays in charge

    Models work on bounded tasks inside your systems, and anything that matters goes to a named person to confirm.

  • Tested on your examples first

    We measure a model on real cases from your work before it reaches staff, so you see where it fails while the stakes are low.

  • We say when AI is the wrong tool

    Not every task should be automated. When conventional software or a person is the better answer, we tell you.

Every AI feature we build follows the same eight commitments, from tested pass criteria to logged decisions. Read our approach to responsible AI.

Benefits

What changes for you.

Illustration for applied ai

  • Less copying and searching

    Staff spend less time re-keying fields or hunting through folders, and more on the judgement the task actually needs.

  • Answers people can check

    Each result links back to its source document, so a reviewer can confirm it instead of taking it on trust.

  • A clear go or no-go

    Testing against your own examples shows whether an approach is good enough before you build a workflow around it.

  • Uncertainty goes to a person

    Review points and confidence checks send doubtful results to someone on your team before they reach your records.

Capabilities

Start with the part you need.

What the work includes

Start with a task someone can describe and check.

We look for bounded work such as finding an answer in approved knowledge or extracting fields from a document. We evaluate representative examples, connect the result to the existing workflow, and decide where a person should review it.

  1. Use-case evaluation

    Test whether AI improves a specific task, using real examples, before building around it.

  2. Knowledge & document workflows

    Find answers in approved knowledge or pull fields from documents, inside the tools you already use.

  3. Human review & safeguards

    Show the source behind each answer and keep a person in charge of anything that matters.

Use cases

Projects we take on.

A practical sequence

How the work runs.

  1. Choose a bounded use case

    Define the input, expected output, permissions and cost of an incorrect result.

  2. Evaluate with examples

    Compare approaches against real task examples for quality, latency and operating cost.

  3. Integrate with oversight

    Add source context, review points and a way to recover when the system is uncertain.

Describe the task ↗

Industries

Applied AI by sector.

What’s changing

Shifts to plan for.

Dates that affect applied ai work, checked against the official sources.

  1. July 2026

    EU high-risk rules move to 2027

    The AI Omnibus entered into force on 27 July 2026, moving obligations for Annex III high-risk systems, such as those used in employment, credit scoring and education, to 2 December 2027.

  2. August 2026

    EU AI transparency duties apply

    Article 50 obligations, such as telling people when they are dealing with an AI system, apply from 2 August 2026.

  3. February 2026

    New UK rules on automated decisions

    The Data (Use and Access) Act 2025 widened when significant decisions can be automated, provided people are told, can contest the outcome and can ask for a human review.

Technologies we use

Tools chosen for the product.

Web
TypeScript
Back end
Node.js
Python
Data
PostgreSQL
MongoDB
Cloud & infrastructure
AWS
Microsoft Azure
Google Cloud

Cost

What shapes the estimate.

  • Document variety

    Many layouts of the same form need more examples and more testing than one consistent template.

  • Cost of a wrong answer

    The more a mistake would matter, the more evaluation, validation and human review the workflow has to include.

  • Data and hosting rules

    Limits on where data may be processed can narrow the choice of models and add hosting work of their own.

  • Where results land

    Output has to reach a system your team uses, and older systems with limited APIs make that connection harder.

Engagement

Ways to work together.

  1. A first, fixed stage

    Discovery, a prototype or a technical review with agreed outputs, so the next decision is made with something concrete in hand.

  2. Delivery in phases

    The build split into releases, each with a scope you review and approve before it starts.

  3. Ongoing engineering

    An agreed scope for support and improvements after launch, with response expectations written down.

Describe the task ↗

Choosing a partner

Questions to ask any provider.

Including us. The answers tell you more than a portfolio does.

  1. How will you measure whether it works?

    Without agreed examples and a pass mark, “it looked good in the demo” ends up as the only test.

  2. Where will our data be sent and stored?

    Model providers differ on hosting regions, retention and training use, and your data protection obligations still apply.

  3. What happens when the model is wrong?

    Every model makes mistakes. What matters is how errors are spotted, corrected and kept out of your records.

  4. Would a simpler approach do the job?

    Rules, ordinary search or a better form sometimes solve the problem at lower running cost. An honest provider will tell you.

Common questions

Before you get in touch.

Do we need to train our own model?

Not necessarily. Existing models, retrieval over your documents and conventional software may cover the task. We assess the simplest approach that meets the quality and data requirements.

Can this work with our existing software?

We assess the available APIs, data formats, permissions and workflow. Integration depends on what your systems expose; we identify those constraints before committing to the build.

How do you handle sensitive information?

We establish what data the task needs, who can access it, and which model or hosting options are acceptable. Data handling requirements and provider choices are agreed for each project.

What happens when AI gets something wrong?

The workflow needs a way to expose uncertainty, review results and recover. We use evaluation examples and explicit checks, with human approval for actions where a mistake would matter.