Research threads that run alongside our products and engagements, newest frontier first. Each is a question working systems keep raising — and a possible collaboration.
Agentic systems · security
Agents that cannot be talked out of their instructions
The question
How do you stop an agent from acting on instructions hidden inside the documents, emails and web pages it reads?
Why it matters
As agents gain tools and permissions, prompt injection stops being a curiosity and becomes an attack surface. The defence has to live in the architecture — permissions, isolation and validation — not in the prompt.
Most relevant to
Banking and financial services, insurance, healthcare, government
Agentic systems · evaluation
Evaluating agents whose path changes every run
The question
How do you regression-test a system that may take a different, equally valid route through a task each time it runs?
Why it matters
Output-only tests miss failures in the steps; step-by-step tests reject valid alternatives. Production agents need evaluation of trajectories, not just of answers.
What should an agent remember across days of work — and how is that memory kept correct, permissioned and able to forget?
Why it matters
Long-horizon work needs continuity, but unbounded memory accumulates errors and leaks access. Agent memory has to be designed like any other stateful system.
Most relevant to
Professional services, legal, customer service, engineering
Conversational intelligence
Speech-native conversation
The question
Can a speech-to-speech model hold context, handle interruptions and stay within policy without a text pipeline in the middle?
Why it matters
Native audio models cut latency and keep tone and timing, but they are harder to inspect, ground and constrain than text. The question is how to keep the control while gaining the speed.
Most relevant to
Contact centres, automotive, healthcare, field service
Validation
Checkable answers from language models
The question
Which parts of a model’s answer can be verified by rules, retrieval or execution — and how should the rest be bounded and labelled?
Why it matters
Enterprise use depends less on how often a model is right than on knowing when it might be wrong. Separating checkable claims from judgment is the practical route to trust.
How should a knowledge system represent what was true when, so that its answers change when the facts do?
Why it matters
Most retrieval treats documents as timeless. Operations, regulation and research are not. Temporal context is the difference between an answer and an out-of-date answer.
How far can distillation and quantisation push a language model before it stops being reliable on a device with no network?
Why it matters
Privacy, latency and connectivity push intelligence onto devices and machines. Knowing where a small model stops being reliable turns that move into a design decision rather than a gamble.
Reading drawings, photographs and manuals together
The question
Can one system interpret an engineering drawing, a field photograph and a technical manual together — and cite all three?
Why it matters
Industrial knowledge is rarely text alone. Grounding an answer across modalities is what turns multimodal models into dependable diagnostic tools.
Most relevant to
Heavy machinery, aerospace, energy and utilities, construction
Spatial computing
Generated worlds for real training
The question
Can generative 3D and world models produce training environments accurate enough to practise real procedures in?
Why it matters
Hand-built 3D is the bottleneck of immersive learning. Generation removes it only if accuracy, scale and safety-critical detail survive the process.
Most relevant to
Manufacturing, energy, aviation, healthcare
Agentic systems
Governed multi-agent execution in low-trust environments
The question
What does it take for autonomous workers to be resumable, bounded and auditable by default — rather than by configuration?
Why it matters
Autonomy is useful only when it can be bounded, resumed after failure and audited afterwards. Those properties have to be defaults of the runtime, not settings someone remembers to turn on.
Most relevant to
Aerospace and defence, government, banking, healthcare
Conversational intelligence
Live conversation as a routing problem
The question
How should a real-time assistant spend a latency budget across models while the conversation is still happening?
Why it matters
A live assistant works inside a fixed window of time. Spending it well across fast and slow models is an architectural problem, not a prompting one.
Most relevant to
Contact centres, sales, recruitment, healthcare
Knowledge systems
Graph traversal and vector similarity for provenance
The question
When does multi-hop structure outperform similarity search for answers that must be traced to their sources?
Why it matters
Similarity finds passages that look relevant; structure finds the chain that makes an answer true. Most real questions need both.
Most relevant to
Legal, life sciences, insurance, engineering
Edge and industrial intelligence
Signal isolation at the machine edge
The question
How small can an acoustic model become and still separate a fault signature from operating noise, on hardware with no network?
Why it matters
On a factory floor or an engine, the signal of interest is small and the noise is large — and there may be no network to fall back on.
With your permission, we use Google Analytics cookies to understand how visitors move through the exhibition — which rooms they enter and which works they open. Nothing is used for advertising, and you can change your mind at any time from the footer.
