Building products that help people think, work, and play.
We design and build human-first AI products that treat the human side as seriously as the technical side. Interfaces that feel conversational and forgiving, systems that explain their reasoning, and technology that fits into how people actually work rather than forcing them to adapt.

Mass
Human insight, reimagined through synthetic minds. A platform that allows you to create synthetic human personas that can think, reason, and respond like real people.

Emmly
Make space for what you imagine. Creative partner that unlocks new ways of thinking, feeling, and making. Emmly empower creators to explore, experiment, and create without limits.


Beryl
E-commerce website for Beryl, a London-based bike light and transport company. Today, they support over 300,000 people using their bikes, e-bikes, e-scooters and cargo bikes to get to work university, and to explore.

Appo
Appo was a startup that built a chatbot for businesses to help them automate customer support and sales for dental practices using AI and chatbots.

AI products built at startup speed, without the agency friction.
Built for AI products
We work with product, data, and engineering teams to ship production AI systems - conversational interfaces, copilots, multi-agent workflows - that teams actually rely on. Working alongside a network of senior consultants and product designers when needed.
Systems > Services
No hand-offs. No queues. Just proven product engineering refined across decades of shipping complex systems for global brands.
Deliver exactly what you need
From RAG integrations to agentic workflows, everything is built with monitoring, guardrails, and proper error handling. Production systems, not prototypes.
Built for founders who need things done right.
RAG & Retrieval
Production-ready retrieval systems for AI products that need to move fast and actually work. Data pipelines, chunking strategies, citation handling, and evaluation harnesses.
Agentic workflows
Multi-agent systems that do boring work end to end. n8n integrations, MCP servers, and human-in-the-loop approvals that help complex AI products feel simple and reliable.
Copilot development
Go from idea to production copilot in weeks. Perfect for AI tools launching internal assistants, customer-facing agents, or task-specific helpers.
Production hardening
We act as your AI engineering team - shipping guardrails, monitoring, evaluation frameworks, and safety checks without slowing you down. Backed by a network of senior consultants and product designers when specialist expertise is required.
Evaluation & Quality
Testing frameworks and regression suites built for startups who need Series-B reliability on a Seed-stage budget.
For over 20 years, I've worked with teams at Google, IBM, Air New Zealand, Kpler, EE, News UK, Tesco, and The Economist, sitting in the space between product strategy and hands-on engineering, designing systems and shipping interfaces that make complex decisions feel simple.
Specialising in building production AI systems - conversational interfaces, copilots, multi-agent workflows, RAG systems, evaluation frameworks, and production observability.
Product design and development
Design and build AI product features from concept to production. We design interfaces that handle uncertainty and AI failures gracefully, build component libraries and design systems, implement user flows for review and control, and develop the APIs and integrations that make features work. Product engineering that thinks strategically about what users need, then builds it.
Vibe code rescue
Stabilise codebases that work but feel fragile. We handle codebase triage and technical debt reduction, add test coverage to prevent regressions, refactor for maintainability, and document the critical paths that matter. The goal is code that works reliably and feels safe to change.
RAG and retrieval
Build retrieval systems that actually work. We design data pipelines for ingestion, implement chunking and embedding strategies that preserve context, handle citation and source attribution properly, and build evaluation harnesses to measure quality. No more retrieval that looks good in demos but fails in production.
Evaluation and quality
Define what good looks like, then measure it. We build evaluation frameworks that test the right things, create regression test suites that catch problems before users do, establish acceptance criteria and quality gates, and run performance benchmarking. Quality becomes something you can measure and improve, not guess at.
Agentic systems and copilots
Agents and copilots that can read, write, and act safely. We design tool-using agent architectures with proper permissioning and access control, build task-specific copilots with clear boundaries, implement review and approval workflows, and establish clear failure paths and containment strategies. Systems that augment human capability, not replace it.
Production readiness
Safety, observability, and automation for systems that run reliably. We implement policy-as-code and guardrails to prevent problems, set up distributed tracing and cost tracking so you can see what happened, build alerting and anomaly detection, and create automated workflows that save time. The engineering that stops you discovering your agent spent £900 overnight or did something it should not.
Tools we use.
We use a range of tools to help us build AI products. Our approach is to pick the right tool for the job rather than standardise on one stack: we evaluate models and platforms on your problem and constraints, then use what fits. That means we are comfortable with multiple providers and runtimes, and we avoid lock-in by designing for swapability where it matters. Below are some of the tools we use day to day.
Your questions answered.
What does Things That actually do?
We build AI products that ship. We work with product, data, and engineering teams to turn pilots into production systems - with monitoring, guardrails, and proper error handling. Architecture, APIs, interfaces, deployment. No handoffs between strategy and code. When specialist expertise is needed, we work alongside a network of senior consultants and product designers.
What is your relationship with other agencies?
Things That is a small, senior product engineering practice. When specialist help is genuinely needed (for example, brand, motion, or deep infra), we work with a network of trusted senior consultants and product designers - always UK-based, or in the same country as our clients. No account managers, no hand-offs, no surprise juniors learning on your budget.
Are you UK-based?
Yes. We're UK-based and work exclusively with UK-based suppliers, or suppliers in the same country as our clients. This ensures data sovereignty, time zone alignment, legal clarity, and simpler collaboration. No cross-border complications, no waiting for support, no currency conversion headaches.
What kind of AI work do you do?
Conversational interfaces, copilots, multi-agent workflows, RAG systems, evaluation frameworks, and the observability that makes them production-ready. The focus is on systems teams actually rely on, not prototypes that look good in demos.
Analysing Christmas Adverts 2025
Mass Vision is a persona-based video sentiment tool: a synthetic screening room for ads. You upload a film, synthetic personas watch it, and narrate reactions moment by moment, explaining why, tied to identity, lived experience, ethics, and context. The output is a tension map (not a safe average): where the same scene splits audiences, who feels seen, and who feels repelled. The Christmas ads analysis shows why this matters: social sentiment gave broad takes after launch; Mass Vision shows the segment mechanics underneath, at specific moments. Run it in minutes, iterate while changes are still cheap, and stop learning the hard way in public.
analysing-christmas-advertsMass
My mum thinks it's cool. My friends think it's as odd as a cat with wings. Getting human opinion is hard. It's time-consuming, expensive, and we're not sure we can trust the result. That's why we built Mass, a platform to deploy hundreds of synthetic human personas that think, reason, and respond like the target audience. They're brutally honest and unfiltered, giving a real sense of what audiences wants in minutes, not months. And at a fraction of the cost. Understand perspectives, validate messaging, explore positioning, and more before committing to the real human research.
www.holdmass.comIf you're building AI products and need help with the human side of the engineering, get in touch at hello@thingsthat.com.
Book a 30-minute call