AI Architect in Financial Services: Lessons from Production

From May 2025 to March 2026, I was the Principal AI Architect at a UK fintech in Edinburgh. The company builds AI agents for UK financial services. This is what I learned.

The context

The company’s product is an AI assistant for financial advisers. The adviser is on a call with a client; the AI agent listens, takes notes, prepares the follow-up, and drafts the compliance report. The stakes are high: UK financial advice is regulated by the FCA, and every piece of advice must be documented, auditable, and compliant.

The challenge: how do you build an AI agent that is helpful enough to be useful, but compliant enough to pass FCA audit?

What I shipped

Four things:

The internal agentic AI platform: the production platform used internally to run AI agents for financial advisers. I built the architecture, the runtime, the safety gates, and the compliance layer. Presented agent guardrails to the FCA sandbox.
Agent guardrails for the FCA sandbox: presented the agent guardrails architecture (safety gates, audit logs, kill switches, runtime policy) to the FCA regulatory sandbox. The guardrails are based on the Substrate Pattern and Defence in Depth frameworks (open source at Neul Labs).
The internal agentic AI platform: the production system that runs the company’s AI agents. Built on the Substrate Pattern (Neul Labs open source) with Defence in Depth (Neul Labs open source). The platform handles agent sessions, tool calls, memory, safety gates, and audit logs.
Presented to the FCA sandbox: the agent guardrails architecture was presented to the UK FCA regulatory sandbox for evaluation. The FCA sandbox is the UK’s framework for testing innovative financial-services products in a controlled environment.

The lessons

1. Compliance is a feature, not a blocker

The banks that won the AI race in 2025-2026 were the ones that built compliance into the runtime from day one. The ones that bolted it on after launch spent 6-12 months in remediation. The cost of building compliance-first is lower than the cost of retrofitting it.

2. The Substrate Pattern works in production

The Substrate Pattern (the four substrates: memory, tool, action, identity) was developed at Neul Labs. I proved it works in a regulated production environment in a regulated production environment. The identity substrate (every action attributed to a specific user with chain of custody) was the single most valuable piece for the compliance team.

3. Kill switches matter more than features

The compliance team cared more about the kill switch than about any feature. “Can you stop the agent in under 30 seconds?” was the first question in every review. The answer was yes — the Defence in Depth model has a manual kill switch reachable in under 30 seconds, plus automated circuit breakers that trip on anomalies.

4. Audit logs are the product

For a financial-services AI system, the audit log is not a byproduct — it is the product. The regulator wants to see: who did what, when, why, and can you reverse it? The audit log answers all four. We spent more engineering effort on the audit log than on any feature.

5. Edinburgh is a serious AI city

Edinburgh has a deep AI ecosystem — the University of Edinburgh’s AI program, the Bayes Centre, and a growing fintech cluster. The company is one of the most interesting AI companies in the UK. The talent is there, the ecosystem is there, and the regulatory environment (UK pro-innovation AI policy) is more workable than the EU AI Act.

What’s next

I left in March 2026 to focus on Neul Labs full-time. The open-source frameworks (Substrate Pattern, Defence in Depth, Tiered Governance Model) are developed at Neul Labs and Skelf Research, not at any employer. I applied these frameworks internally to the internal agentic AI platform and presented the agent guardrails to the FCA sandbox. The frameworks are the foundation of the Neul Labs agent infrastructure and the consulting practice at dipankar.co. The consulting practice at dipankar.co helps other financial-services firms implement the same patterns.

— Dipankar Sarkar, Founder of Neul Labs