AI for Finance (Lab)
Agents that can answer to a regulator
The hard parts of applied AI in finance: grounding, refusal, evaluation tied to the firm's own record of truth, and making a run legible when it fails.
Working builds
Systems built end to end and measured. Each result is reported as it came out, failures included.
Working build
Tax Policy Navigator
Right behaviour on 48 of 50 UK tax questions, yet only 195 of 261 cited claims, 75 percent, were fully supported by the paragraph cited.
What it takes to make a generative answer about UK tax trustworthy enough that a citizen could act on it.
Working build
Rule Grounding
Every FCA Handbook citation an agent makes, checked by three tests that use no model. The check an audit needs cannot be built yet: the API serves only today's rulebook.
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.
Working build
Governed Data Eval
Same model, same question: £2,512,500 on the raw systems, £1,512,500 on governed data. The gap is one superseded account.
A wealth-management demonstration of why agent reliability depends on governed data and deterministic, firm-owned evals.
Design prototypes
Clickable interfaces with no model behind them, built to answer a product question before anyone writes the backend.
Design prototype
AgentScope
What should a debugger show when one agent spawns five, and one of them fails?
Two interface prototypes for a developer-facing harness. AgentScope makes a multi-agent run legible at a glance, including when it fails. Its second view, ContextScope, shows how the agent's belief changes from turn to turn.
Design prototype
Data Discovery Agent
What must a data-catalogue agent show next to every recommendation, when the wrong dataset means a wrong answer, a compliance breach or runaway cost?
An interface prototype for a financial-data catalogue where every recommendation arrives with its lineage, entitlement status, and monthly cost.