About Projects Clients News Contact

AI that works offline. And in the real world.

Tools that analyse and automate. Systems that understand movement. Characters you can talk to. Software that holds up in the physical world. Offline, wherever it can be.

Reflex Arc is a UK AI software and app development studio, based in Leeds and working with clients across the UK and abroad. “AI” is one of the tools we use, where it has a purpose, and a person always makes the final call. It goes into tools people use every week, experiences people step into and talk to, and software that runs on devices and machines, not just screens. Wherever we can, it runs offline: on the phone, the headset or a computer in the building, so it keeps working without the internet and the data stays where it is.

In short

Offline first. Models that run without the internet, on the device or on a computer in the building (edge AI, if you prefer the term), now range from ones small enough to live on a phone, to ones that run on an ordinary desktop computer, to much larger ones that need a server of their own. We use the first two every day, and we know when a project calls for the third. Online models still have their place, and plenty of requests have to go to them for now; the aim is that fewer do over time. For a client that means three things: the product doesn't depend entirely on someone else's service staying up, the data stays where it is, and the running cost of paying per request, the bill that grows as a product succeeds, stays under your control.

Much of our recent work is under NDA, so this page describes it by shape rather than by client: the kinds of systems we build, the technologies involved and some of the engineering problems we solve.

Who this is for, and who it isn't

Businesses with a repetitive, document-heavy process that a tool could take off people's hands, with a person still signing off. Studios and agencies who need a partner for the AI, physical or embodied part of a project; we're used to working white-label inside other teams' pipelines. Museums, health and education organisations with an audience who'd benefit from something they can talk to. And anyone with a question about whether “AI” belongs in their project at all, which is a consultancy conversation before it's a build.

Where we'd point you elsewhere. Automating processes across a whole enterprise, customer-service chatbots as a product in their own right, data science and forecasting, and training or running large models at scale are better served by firms built for that, and we'll say so. The same goes for generated images and video. What suits us is a defined piece of software with AI inside it: a tool, an app, an installation or an experience, built by a small senior team you deal with directly, wherever you are.

What we build

AI-powered tools and workflow automation

Practical systems built around the way people already work.

We build practical tools around existing workflows: systems that read documents, organise information, connect services and prepare decisions for a person to review. The aim is to remove repetitive work without hiding what the system has done. That includes analytics: making sense of how people actually use a product or experience, so the next version is better.

A worked example: our own bookkeeping. VATCheck is an internal tool we built for ourselves and use every week. It reads receipts and invoices out of our email, has a language model pull out the supplier, date, amounts and VAT, and matches each one against the bank, attaching the receipt only when everything agrees and leaving the rest for a person. Nothing is written without an explicit go, every action is logged, and the extraction can run on an online model or on a local one that never leaves the machine: the same choice we offer clients. It turned a quarter's VAT reconciliation from days into an afternoon.

An R&D project: home control. A Mac mini in a family home runs the TV, music, lights, games consoles and router, from iPhone and Apple Watch apps we built, a web app and Siri. Simple requests never leave the phone. Requests that need wider knowledge go to an online model, with a model in the house planning the same request alongside it; once the local one has shown it gets a kind of request right, that kind stays home. Where a model is involved it only plans: ordinary code does the work and checks it. Read about our home control system for Sonos, Hue and Apple Watch.

The same approach applies to client projects: understand the existing process, automate the repetitive parts, integrate with the systems already in use and keep people in control of the decisions that matter.

Vision, media analysis and movement

Recognising signs, tracking bodies and reading large collections of media.

Our first shipped machine-learning system was the Makaton signing game in 2015, with a recogniser trained on Makaton signs. It was BAFTA-nominated and runner-up at NHK's Japan Prize. Since then, our work with bodies and movement has also used camera-based pose estimation, skeletal tracking and conventional real-time interaction systems: Rehab Trainer turns shoulder rehabilitation into games patients want to repeat, and more recently a fitness game ran entirely on a phone (one camera, on-device pose tracking, no headset, the skeleton in the tile on our homepage), which we worked on through its 2024 private beta. Not every part needs machine learning: the point is choosing the right technique for the interaction.

The Makaton signing game, played by performing signs in front of a Kinect

The Makaton signing game (2014–15), played entirely by signing, with a recogniser trained on Makaton signs. Developed with GamelabUK and Hassell Inclusion.

Body tracking runs through our accessibility work too: for Sky, The Rising: Escape the Dark used Sky Live's camera-based body tracking in the living room, designed so blind, partially sighted, deaf and hard-of-hearing players can play alongside everyone else.

A more recent strand: pipelines that run vision and language models over large collections of media and turn what they find into structured results a person can review, query and act on, work that would otherwise be done by hand, one item at a time. Built to run in batches at scale, with the models' output checked against human judgement before anyone trusts it. The questions behind it are ones we first worked on prototyping curation tools for BBC R&D; the models have caught up since.

Conversational AI

Speech in, speech out, a character or companion that answers in the moment, in a headset, an app or a room.

A person speaks; a character listens, thinks and replies in a voice that fits, fast enough that the conversation feels like one. Connecting a language model to a voice API is the easy part. We build the complete experience around it: the real-time application, interaction design, animation, safeguards and system integrations that make the conversation usable: speech recognition, the language model that decides what to say, the guardrails around it, speech synthesis, and the lip-sync, gaze and body language that keep the character in the world rather than floating over it.

We have extensive experience with ElevenLabs for real-time voice, usually combined with a language model from Anthropic (Claude), OpenAI or Google for the conversation itself, chosen per project, not by habit. The difficult parts are latency you don't notice, turn-taking that doesn't talk over people, knowing what the character must never say, and a graceful hand-off to a human when that's the right thing.

The shape of it, from projects we can't name: a voice-led companion built to be talked to in confidence, designed so that sensitive processing stays on the device where that's possible. Assistants that talk visitors through an exhibit. Characters in a headset who remember what you told them ten minutes ago. We've built several of these, all under NDA, and we're happy to talk through how they work on a call.

One we can name: for Riptide: Intermission (2023), an 80-minute mixed-reality meditation with XR Stories, each visitor began with an AI-led onboarding interview on Charisma's platform. Charisma built the conversation; we built the VR experience that listened to it, taking what each person had said and shaping how their journey looked and behaved, so no two were the same.

A visitor wearing a headset during Riptide: Intermission

Riptide: Intermission (2023), with XR Stories. Charisma built the AI onboarding conversation; we built the VR experience that responded to it.

Before all of that: in 2017, for Channel 4's The Robot Will See You Now, we built the emotive face for Jess, a robot that sat in a room with real couples and families: idle, listening, thinking, reacting, on an Android device.

Physical AI: REK’s robots and simulator

Simulation and real-time control, with a person driving.

REK connects game-engine simulation with real humanoid robots, and we're part of the team building its multiplayer simulator, which lets people train and qualify remotely before controlling physical robots. We also worked on the VR controls for REK's first event, REK0. The simulator runs the same motion policies as the real robots, and REK's real-time systems connect the human pilot to the machine.

This is human-directed control rather than autonomous decision-making: the trained systems handle balance and movement, while a person decides where the robot goes and what it does. It's taught us what it takes for software to hold up in the physical world, in front of a live audience, with no second take.

Same controls, real and virtual: a REK robot driven from a handheld controller in REK's workshop (top), and REK Sim on a Steam Deck (bottom). We're part of the wider REK team, helping build the simulator and its pilot controls. Video: Cix Liv / REK.

AI consultancy, feasibility and prototypes

Before the build, or instead of it.

Sometimes the useful work happens before a build. We can assess a brief or existing implementation, test whether an idea is technically viable, produce an architecture and delivery plan, or build a focused prototype to resolve the important unknowns.

An engagement might be a feasibility assessment, a technical architecture review, an independent second opinion, a prototype, or ongoing technical guidance across a project. We'll cover likely costs, latency, privacy, deployment options and the parts that don't need AI. We've done this as immersive advisors on a three-year Wellcome Trust research programme (LivingBodiesObjects) and as the technical second opinion inside agencies and studios.

You get a plan you can build with anyone, including us.

How we think about it

We'll tell you when it's the wrong tool. Some of the most useful conversations we have end with a simpler answer than the brief asked for.

Accessible by design. Voice, body tracking and spatial audio are also the technologies that open experiences to people a screen and controller shut out. We design for that from the start. It's why the accessibility team at HSBC commissioned us, not the marketing team.

No AI-generated images or video. Making pictures, footage or artwork with generative “AI” isn't what we do, and we don't offer it. We'd rather work with the artists, photographers and filmmakers who make that work than try to replace them. Where it helps them, AI can sit alongside the software they already use, connected through MCP or something more custom, to speed up the repetitive tasks, while they keep complete creative control.

Questions we get asked

Can it run on-device, without sending data to the cloud?
Often, yes. We look at what can run on the device or on a computer in the building, what has to go to an online model, and what each costs in latency, quality and money, before you commit. That includes native Android; Meta Quest headsets are Android underneath.
Do you build customer-service chatbots or call-centre voice agents?
Not as a rule. Our work is characters, companions and assistants inside experiences: VR, mixed reality, installations, health and learning apps, live events. If you need a support bot for a website, there are specialists who do that better than we would.
Can we see examples of your voice AI work?
Most of it is under NDA, so we can't show it publicly. On a call we can talk through how those projects were built, the problems we hit and how we solved them, without naming clients.
What does working with you look like?
Most projects start small: a call, then a feasibility assessment, second opinion or focused prototype that answers the important unknowns before anyone commits to a full build. We agree up front who owns what, and how any data will be handled: where it's processed, where it's stored and who can see it.
Will you tell us if AI is the wrong tool?
Yes. We'd rather build the right thing than the fashionable one.

Something in mind?

A process that eats hours every week, a character people could talk to, a companion inside an app, a tool that needs to understand a body, a collection of media that needs reading, or just a question about whether “AI” belongs in your project at all. We consult as well as build: tell us a little about it and we'll come back to you.