ArchiSoft.

UK AI research & development

Useful intelligence.
Examined carefully.

Exploring how AI can help people understand complex systems, evaluate architectural options and carry out bounded tasks with accountable oversight.

Research direction · Public overview

The questions shaping our work

Evidence before
confidence.

ArchiSoft is a UK-based consultancy. These themes set out our research direction; they do not represent published results or a claim of programme eligibility.

01 / Decision support

Can a recommendation show its working?

Architectural decisions draw on requirements, operational experience and incomplete documentation. Plausible advice can overlook a critical dependency.

  • Investigate: Methods that link recommendations to source evidence and expose uncertainty.
  • Evaluate: Requirement coverage, factual support, contradictory advice and verification effort against manual and simpler AI baselines.
  • Potential use: Help architects prepare and review decisions while retaining final judgement.
02 / Evolutionary architecture

Can we see the consequences of change sooner?

Code, architecture records and operational behaviour do not always tell the same story.

  • Investigate: AI-assisted comparison of proposed changes with recorded architectural constraints.
  • Evaluate: Missed violations, false positives, review time and useful suggested checks on controlled changes.
  • Potential use: Earlier feedback about dependencies and architectural rules as systems evolve.
03 / Trust & assurance

When should an agent stop and ask?

An agent with tools needs explicit boundaries, especially when an action has meaningful consequences.

  • Investigate: Combinations of permissions, approval points and runtime checks for bounded engineering tasks.
  • Evaluate: Task completion, unauthorised actions, appropriate escalation and recovery under misleading instructions and tool failures.
  • Potential use: Carefully bounded agents with accountable human oversight.
04 / Compute efficiency

How much model capability does this task need?

A larger model may add cost and delay without improving a particular result enough to justify them.

  • Investigate: Task routing, retrieval, caching and smaller models against defined quality thresholds.
  • Evaluate: Accepted results, latency, inference cost and directly measured compute use; keep estimates distinct.
  • Potential use: AI architectures selected for quality, operating cost and maintainability.

Featured product initiative

AI Company Brain

A practical focus for questions about organisational knowledge, evidence and decision support.

The public product overview describes the ambition and potential uses. Proprietary implementation details and confidential information remain private.

Explore Company Brain →
Public research record

Make the evidence inspectable.

A credible research update should identify the UK work undertaken, the technical uncertainty, the method and the result.

  • Named research lead, UK location and activity dates.
  • Prototype or method and its baseline.
  • Evaluation scope, conditions and measured findings.
  • Limitations, next experiment and shareable evidence.

Detailed research records are not yet published on this site.

Collaboration in the United Kingdom

Bring a question
worth investigating.

We welcome conversations with UK businesses, universities and research organisations.

A collaboration could begin with a representative evaluation task, an assurance question or a prototype needing independent assessment. Our strongest interests are trust, human oversight, compute efficiency and software productivity.

Discuss research collaboration ↗

A useful conversation starts here

What is getting
in your way?

Bring the difficult decision, the slow-moving system or the AI idea.
We’ll start by understanding your situation.

Discuss your challenge