Field Notes
Practical, opinionated articles from DTA: agentic engineering, on-device AI, architecture and transformation in regulated organisations.
Field Notes preserve the practical Insights voice. The Observatory is where those signals become living models.
June 22, 2026 · payments
The Bank of England has published its Policy Statement on sterling-denominated systemic stablecoins, and with it, in Appendix 4, a draft Code of Practice for systemic stablecoin issuers. The Bank intends to finalise the Code by end-2026, after which it...
June 18, 2026 · Banking, Regulation
The FCA has stopped asking whether banks implemented the Consumer Duty and started asking whether customers are genuinely better off. That turns the Duty into a loop: test before launch, learn from outcomes, evidence both. Two shifts now arriving, agents...
June 18, 2026 · Banking, Architecture
Three years into the Consumer Duty, the FCA keeps finding the same gap: firms can show what they did, but not whether it improved the outcome for the customer. That is not a reporting problem. It is an architecture problem,...
June 17, 2026 · Banking, Architecture
What if a branch device could answer a customer's question with no network round-trip — and cryptographically detect whether its own records have been tampered with, offline, without trusting the centre? That is edge CQRS with a hash-chained ledger. Here...
June 17, 2026 · AI, Architecture
The reflex for agent memory is a vector store. In a regulated bank that is the wrong default. An agent's memory should be a governed projection of an event log — so it can be replayed, audited, and scoped by...
June 17, 2026 · Banking, Architecture
Most banks choose a platform first and discover the hard parts later. The independent move is to define a target architecture and prove the primitives that usually fail before committing, so the platform conforms to the architecture rather than the...
June 16, 2026 · Banking, Architecture
Regulators moved operational resilience from a reporting exercise to a demonstrable capability. Most banks answered with documents and runbooks. The real answer is architectural: a platform that watches itself, diagnoses its own incidents, and can tell you — fast —...
June 16, 2026 · Banking, Architecture
A transaction foundation model turns a bank's own event history into one reusable intelligence layer. But a model trained on raw events is not operable in a regulated bank until the data beneath it is governed. The missing layer is...
June 15, 2026 · Strategy, Architecture
Agents are crossing from advice to action. A bank can only allow that if it stays accountable for every action and sovereign over the agents, data and models involved. We call that condition Owned Autonomy, and the C3 architecture spine...
June 15, 2026 · Strategy, Architecture
Consent is the control plane of the C3 spine: authority, policy, evidence and an agentic gateway. It's how a regulated bank lets an agent do more than advise while keeping explicit, inspectable control of what it's allowed to do —...
June 15, 2026 · Strategy, Architecture
Context is the quiet plane of the C3 spine — a shared semantic layer plus measurement. Skip it and your capabilities won't compose, your agents can't be trusted across domains, and you can't prove the economics. Meaning is architecture, not...
June 14, 2026 · Strategy, Architecture
Owned Autonomy needs an architecture. A bank can only be accountable for what its agents do, and sovereign over what they run on, if three things are shared across every proposition: governed capability, shared meaning, and the authority to act....
June 14, 2026 · Strategy, Payments
Cross-border payments are being rebuilt on real-time rails and pre-funded local accounts. Most of the conversation is about cheaper, faster payments. The bigger prize is the deposit — and the net interest, FX and liquidity revenue that sits on top...
June 13, 2026 · Banking, Architecture
Most banks already have a capability map. BIAN tells you what the capabilities are. It does not tell you how to package and run one so it can be reused, governed, and proven. That missing unit is the Domain Building...
April 11, 2026 · Engineering, Operations
A six-hour hang and a fake embed bug, both caused by the same thing: a loop that had no way to prove it was still working. Two fixes, one rule.
April 8, 2026 · Engineering, AI
How I built a ML pipeline to convert 5,287 voice memos recorded over five years of banking consulting into a structured book on UK digital transformation—using local models, immutable lineage, and staged quality gates instead of one-shot prompts.
April 8, 2026 · Engineering, AI
How to build a derived graph layer on top of markdown files so you can ask 'what contradicts this?' instead of just 'does this exist?'
March 26, 2026 · Engineering, AI
An article on agent-friendly CLIs went around recently. It covered the basics well but missed the patterns that matter most. We built a Claude Code skill that fills the gaps — structured output, stderr separation, exit codes, TTY detection —...
March 26, 2026 · Engineering, AI
Writing a Claude Code skill is easy. Knowing whether it actually works is harder. We built an eval-and-tune loop that benchmarks skills against a no-skill baseline, and the first thing we learned is that your initial assertions will be wrong....
March 23, 2026 · Engineering, Security
Vapor is a capable web framework. But if you want enterprise-grade multi-tenant security — OIDC, RBAC, audit trails, tenant isolation — you're writing it from scratch. We built the chassis that closes those gaps, and the toolchain that made it...
March 20, 2026 · AI, Engineering
Cloud AI APIs are powerful, but some data can't leave the building. We're building a pipeline that trains small, domain-specific language models to run entirely on-device — no API calls, no data exfiltration, no per-token costs. Here's why, and how....
March 20, 2026 · AI, Strategy
A client needed 50 PowerPoint decks converted to a new brand identity. Our first instinct was to train a model. We were wrong. Here's the pipeline we built instead, and the decision framework that stopped us from over-engineering.
March 18, 2026 · AI, Apple
Hands-on benchmarks comparing Qwen3.5 2B, Qwen3.5 4B, and Llama 3.2 3B running locally on Apple Silicon via MLX-LM. Practical findings on speed, memory, output quality, and the sampling parameters that actually matter.
Tell us the proposition, architecture decision, or transformation problem you are facing. We will tell you, honestly, whether and how we can help.
Discuss a challengeLiving models and Field Notes on the future of banking. Published when the model changes, not on a schedule.