AI Lab
Experiments and prototypes.
A working portfolio of small, opinionated explorations in what an enterprise AI product needs before a regulated institution can act on its output: the data agents draw on, the tools they act through, and the limits around them. Each one starts from a single product question and tests one of the controls that decides whether the answer can be trusted: refusal, evaluation, observability, entitlements, lineage, cost. Each is built as the cheapest artefact that makes the answer legible, each carries a writeup of what it was testing and what fell out, and most have a clickable demo.
-
Prototype
Rule Grounding
An agent reviews a financial promotion and cites the rules it applied. Every citation is then checked against the regulator's own rulebook by three checks that involve no model at all.
A working prototype of citation grounding for regulated agent output, built against the FCA Handbook API on the day after it launched. It resolves a citation, checks whether the provision is binding, and refuses to certify a decision whose governing text has changed since it was taken.
What it established The check worth building is not whether the agent understood the rule. It is whether the provision exists, and whether the regulator publishes it as a rule at all rather than as guidance, which binds nobody. Both are string comparisons against a field the agent did not supply. The check that cannot be built is the one an audit actually asks, because the API serves today's rulebook and refuses the version that applied on the day the decision was taken.
Interface prototype · Agent governance · Agent-ready data · Evaluation · Financial services
-
Prototype
ContextScope
AgentScope for agent belief state. An interface prototype that shows the conversation on the left and the agent's changing case state, as a turn-by-turn diff, on the right.
A clickable prototype for a causal debugger of agent state: what the agent believed, why it believed it, what conflicted, and what that enabled next. Where AgentScope makes a run legible, ContextScope makes the agent's belief legible. Built around the conflict a regulated operation has to handle well: the ledger says the card payment was authorised, the customer says it was not them.
What it established A transcript is too thin to debug belief and a full state snapshot is too large to read every turn, so the unit that works is the diff between turns. That one choice is what turns a context conflict, the ledger says the payment was authorised and the customer says it was not them, into a visible event rather than something reconstructed afterwards.
Interface prototype · Product design · Agent observability · Governance
-
Case study
Tax Policy Navigator
What it takes to make a generative answer about UK tax trustworthy enough that a citizen could act on it.
A case study in product judgment for grounded AI in a regulated domain. Why refusal has to be a designed surface, why a citation has to reach down to the individual claim, and where a machine judge stops being enough and a human has to take over. Built and evaluated end to end over the full HMRC Employment Income Manual. The hard problems turned out to be product problems, not model problems.
What it established Over fifty questions on the HMRC Employment Income Manual it did the right thing 48 times, and only 75 percent of its individual cited claims were fully carried by the paragraph they pointed at. Answer-level accuracy hides claim-level reaching, which is why the citation has to reach down to the sentence.
AI product · Grounded generation · Evaluation · Regulated domains · Trust and safety
-
Prototype
AgentScope
An interface prototype for a developer-facing harness that makes a multi-agent run legible at a glance, including when it fails.
A clickable prototype that visualises a tree of agent runs, each with their own steps, context window, and cost share, on one screen. Three switchable mock runs demonstrate the design across different shapes.
What it established The readable unit of a multi-agent run is the tree over time, not the transcript. Once children indent under their parent and parallel agents stack on the same time range, the sub-agent that burned the budget is visible without opening a single span.
Interface prototype · Developer tools · Multi-agent · Observability
-
Prototype
Data Discovery Agent
An interface prototype for a financial-data catalogue where every recommendation arrives with its lineage, entitlement status, and monthly cost.
A clickable prototype that answers a question most chat-with-data demos refuse to ask. How should a regulated-data platform expose an agent surface when picking the wrong dataset can mean wrong answers, compliance breaches, or runaway cost?
What it established Lineage, entitlement status and monthly cost cannot live in three other applications. Rendered anywhere other than next to the recommendation itself, at the moment of choosing, nobody consults them.
Interface prototype · Product design · Data catalogues · Compliance
Also
-
The reasoning was defensible. The answer was wrong. Prototype
A runnable wealth-management demonstration of why agent reliability depends on governed data and deterministic, firm-owned evals.
Same model, same question, two answers: 2,512,500 pounds on the raw systems and 1,512,500 on the arranged view. The gap is exactly one superseded account, and no external benchmark could have caught it, because which account supersedes which is a decision the firm makes rather than a fact a model can read.