SR3H

SR3H · Oxford

practical intelligence, engineered.

SR3H builds useful software and researches the systems that help language models work reliably with people over time.

We design the layer around the model — interfaces, structured context, scoped memory, persistent state, permissions and evidence — so capability remains coherent, accountable and under human control.

Explore our work Explore our research

Independent product and research studio · based at Grassroots, H B Allen Centre, Oxford

Selected work

Products with a clear reason to exist.

Each starts with a practical problem and a clear job for the person using it.

Conchup identity mark

Persistent engineering context

Conchup

Requirements, decisions, implementation, tests and evidence connected across long-running human–AI development.

Brane identity poster reading Your context. Under your control.

Context and privacy

Brane

A local-first context system for giving AI relevant personal knowledge while keeping memory, roles and permissions under the person’s control.

Proper Job home screen

Construction pricing

Proper Job

Clear guide-price ranges for UK building work, showing the assumptions and cost drivers behind each estimate.

NextUp welcome screen

Viewing tracker

NextUp

A personal viewing log for keeping track and deciding what to watch next.

Illustrated landscape for Drift

Local discovery

Drift

A place-discovery tool for exploring nearby destinations and deciding where to go next.

Good Route route-selection home screen

Route navigation

Good Route

Navigation that preserves the rider’s chosen route and makes rerouting an explicit decision.

Redthread private briefings sign-in screen

Private briefings

Redthread

A private workspace for developing briefings, with each person’s work kept separate.

Human–AI systems engineering

Making capable models work in real systems.

A capable model is only one part of a dependable system. We work on the surrounding layer: capturing intent, selecting relevant context, scoping memory, authorising tools, preserving state and producing evidence for what happened.

The goal is to turn model capability into coherent, accountable work over time — without obscuring human judgement or exposing information the task does not require.

  1. Human intent
  2. Scoped context
  3. Model + authorised tools
  4. Persistent state
  5. Evidence
  6. Human control

Research systems

Persistent context. Deliberate boundaries.

Conchup and Brane explore two sides of the same problem: how useful context can persist across complex work while remaining bounded, inspectable and under human control.

Persistence for the work

Conchup

Conchup maintains persistent engineering context across requirements, decisions, design, implementation, tests and release evidence.

It is designed to trace the impact of a change through affected work, preserving what changed, why it changed and the evidence that the resulting system remains correct.

Discuss Conchup

Context controlled by the person

Brane

Brane explores how people can provide AI with relevant, persistent context while keeping data, roles and permissions under their control.

Rather than one undifferentiated memory, work, private life and other contexts can remain deliberately separate.

What does this system need to know for this task — and what does it not need to know?

Lunchtime Websites

Small, clear websites for people who need to be found online.

A small pro bono SR3H initiative for people and good causes that need a clear, useful website.

Anup House in Bodh Gaya Lunchtime Websites · Web Anup House Bodh Gaya, India

The first example: a clear online home for an independent guesthouse in Bodh Gaya, improving visitor information and discoverability.

Know a person or good cause that needs a simple website? Suggest a project.

How we work

Quietly capable.

  • Start with the real problem. Build the smallest coherent thing that materially improves it.
  • Make every layer agree. Language, interaction, design and engineering should describe the same product.
  • Keep context relevant and boundaries explicit. More information is not automatically better information.
  • Preserve human judgement and show evidence. Important decisions should remain understandable and checkable.

Work with SR3H

Have a useful problem, a serious idea or a reason to build?

We welcome thoughtful conversations with potential partners, customers, collaborators and people interested in the work we are growing.

hello@sr3h.uk