An applied AI lab, where ideas get built.

How far can a model act on its own, and where should it stop?

A model that answers is useful. A model that plans, uses tools and finishes a task is something else: it needs boundaries, memory and a way to show what it did. Most of the work is in the harness around the model, not in the model itself.

• Tool calling against real apps and APIs
• Harnesses: permissions, guardrails, confirmations, logs
• Long-term memory, with a human in the loop for anything that matters

What changes when there’s no screen to fall back on?

Voice is the most natural interface, and the least forgiving. Latency, interruptions, accents and hesitations all count, and there’s no button to tap when the system gets it wrong. It’s also what opens AI to people who don’t use screens easily.

• Real-time speech recognition and synthesis
• Turn-taking, interruptions and sub-second responses
• Conversations that keep their context over days, not one exchange

Can you get the benefit of large models without giving up your data?

For many organisations, sending data to a cloud they don’t control is out of the question. Open-weight models have closed much of the gap with the big closed ones; the question now is how to run them well where the data already lives.

• Open-weight models on private servers (on-premise)
• Compact, quantised models running on the device itself (on-device)
• Hybrid set-ups that keep sensitive data local and pick the right model for each task

What does AI look like when it works for a group, not one person?

Most assistants are built for one user and one prompt. Real life happens in families, teams and shared lists: several people, each with their own role, and things that change while you’re not looking.

• Shared spaces where several people and an agent work together
• Roles, permissions and a clear record of who asked for what
• Keeping everyone informed without flooding them