Changing regulated platforms
Retire systems, move data and rotate credentials without losing the records or the audit trail.
Writing
Notes from the build
Architecture, delivery and lessons learned, from ingesting the Twitter firehose for trading firms to building white-label banking platforms and practising agentic engineering.
Start here
Follow a thread
A few routes through the writing, selected by problem rather than publication date.
Payments and commercial rules
Make payment timing, prices and discounts explicit contracts rather than scattered configuration.
Engineering with agents
Agree the contract, challenge the tests and keep the author separate from the reviewer.
Platform judgement
Count migration effort and opportunity cost, then fix the bottleneck that matters.
Archive
All writing
2026
Mobile Was Never the Hard Part: Enterprise Mobility Was Integration
I once sold a WorkSite mobility product to Interwoven. Looking back, the interesting part was not the handset UI but the server-side integration, queuing and extension points that made mobile workflows dependable.
Platform Limits Shape Products More Than Ideas: Lessons from a Pre-AirTag Child Tracker
Before AirTag, we built a prototype child-tracking system around Bluetooth tokens, an iPhone app, shopping centre receivers, and a Rails backend. Looking back, the strongest lesson was that platform rules, hardware economics, and privacy constraints mattered more than the code.
Good Engineering Habits Age Better Than Technology: Rereading My Final-Year Project
Rereading my final-year project from the mid-1990s, I was struck by how little the core lessons have changed. The stack is obsolete, but the habits still look right: understand users, iterate, separate concerns, write things down, and design for change.
Modelling Batch Pharmacy Claims in FHIR
FHIR has no ClaimBatch resource, and NCPDP batch files don't care about REST. How we composed Task, List, Claim, ClaimResponse, ExplanationOfBenefit and PaymentReconciliation into a claims lifecycle with at-least-once delivery that's safe to retry.
Right-Size Every Token: The Pi Utilities That Keep an Agent Workflow Affordable
Token pruning, native REST tools that skip the model, a three-tier agent budget, and the small terminal extensions (vi mode, statusline, notify, theme) that make a Pi workflow comfortable all day.
Skills Are Commands, Agents Are Specialists: How My Pi Day Is Organised
The six natural-language skills, the fleet of phase and review agents, and the MCP servers that sit behind my Pi workflow, from the morning briefing to commit and pull request.
Transparent Hooks Beat Explicit Orchestration: Inside My Pi Dev Harness
How a Pi extension runs a full agentic SDLC through hooks the agent never sees: shift-left gates by phase, TDD with isolated test authoring, a dangerous-command gate, plan mode and git worktrees.
The Author Is Compromised: Independent Review and an Honest List of Gaps
Independent review agents, decision packets built for thirty seconds, and a plain list of what ACCORD has cut or only half built. How to keep an agent harness honest about its own limits.
Write the Tests Before the Code, Then Review Them Like an Adversary
Why ACCORD pushes decisions left: tests written from the spec alone, an adversarial review before any production code exists, and three layers of protection that keep most checks off your attention.
Medley: One Command to Run the Whole Platform Locally
A seed job, a build artifact, five services across three package managers and an AWS emulator, all brought up in dependency order and wired to each other by one command. How we built a thin orchestration layer from a YAML registry, a pure-JS resolver and Tilt, and the things we deliberately didn't build.
How I Work Today: Terminal, Agents and a Second Brain
Four years on from my Babylon-era setup: agent skills in Pi, tmux and Neovim as the whole IDE, Gmail in the editor, dotfiles that build a machine in one command, a PARA notes repo as a second brain, and the same session from desk, iPad or iPhone over Tailscale and mosh.
The Cheaper Platform Isn't the Cheaper Migration
A replacement workflow engine would have cut the licence bill by more than 90%. The honest business case said the question wasn't cost at all. How to write a platform-migration case that counts engineering time, opportunity cost and operational maturity honestly.
Data Residency Without a Full In-Country Stack
Entering a market that requires patient data to stay in-country doesn't have to mean cloning your whole platform. Split it into data cells and processing hubs, then hunt down the places data quietly comes to rest.
I Rebuilt My Agent Pipeline Three Times. The Substrate Was the Problem.
Most agent setups optimise for capability. I optimised for attention, and built ACCORD, a Pi extension that turns a free-text request into a verifiable contract before anyone writes code.
Fix Developer Experience Before You Re-Platform
When one feature means five coordinated PRs and a shared dev environment, a new runtime platform won't save you. A monorepo, a one-command local stack, contract tests and preview environments will, whether the humans or the AI agents are writing the code.
One List Price, Many Coupons: A Pricing Standard for Stripe
Every promotion modelled as a new Stripe price multiplies into hundreds of IDs that drift. Here's the standard we adopted: prices are list prices, every adjustment is a coupon, and plans are versioned documents, not dashboard clicks.
Search-Then-Create Is a Race Condition
Conditional create on top of an eventually consistent search index will create duplicates under load. A small, strongly consistent write barrier fixes it, and the hard part is choosing the uniqueness key, not the lock.
Rotating Credentials Without Downtime (and Knowing What They Touched)
A rotation playbook for third-party credentials across environments and regions: overlapping validity, dual-running webhook secrets, a boring checklist, and the audit trail you need before an incident, not after.
Pause Is Not a Status: Modelling Holds as Overlays
A requirement to 'pause' a care programme turned out to be four independent behaviours. Model it as a time-bounded hold with declarative effects rather than a lifecycle status, and one mechanism covers every kind of delay.
Bringing E-Signing In-House Without Prescribers Noticing
A per-signature SaaS bill that grows with every prescription. Designing an in-house Advanced Electronic Signature pipeline to replace it: the cryptography is the easy part; the migration design is what makes it safe.
Building an AI-Native Development Workflow with Pi
Eleven extensions, six skills, eighteen agents and a token budget: how I rebuilt my whole working day inside a terminal-first coding agent, from morning briefing to verified pull request.
Evaluating a PaaS That Runs in Your Own AWS Account
We wanted a Heroku-like developer experience without leaving AWS or re-platforming our data. A requirements-first evaluation of managed runtimes found two gaps every vendor shared, and a few questions worth asking before any pilot.
A Spec-Driven Agentic Workflow With Claude Code Skills
Five skills, thirteen single-purpose agents and a spec that travels with the code: how I structured AI-assisted delivery so every session starts fresh and every change is verified against written acceptance criteria.
Load Testing Told Us Autoscaling Was Unsafe
Adding pods made things worse. Before you switch on a Horizontal Pod Autoscaler, you need admission control, bulkheads and fan-out budgets. Otherwise more replicas just means more retries hitting the same bottleneck.
2025
Write the Rules Down First: Discount Stacking and Payment Accounts
Two payments RFCs from the same week: a stacking policy for employer coverage, coupons and referral credits, and a design for several payment accounts in one deployment. Both are cheaper to decide before marketing and partnerships need them.
Preparing for a TV Campaign: Fix Authorisation First
A TV advert means a traffic spike you can't schedule precisely, because catch-up viewing exists. Our plan: config-only scaling first, fix the one call every request makes, then add application-level protection with a go/no-go gate two weeks out.
Make the Invoice Service Dumber
A payment service that knows about referral credits, employer coverage, promotions and shipping has to change for every new business idea. Push business rules out into resolvers that produce line items, and the core never changes again.
An Eligibility File Is a Contract
B2B health programmes live or die on two files: who is eligible (in) and what happened (out). Design both as versioned contracts with clear keys, and you get easier onboarding, easier reporting and fewer awkward client calls.
When 'Shared Nothing' Means Duplicating Everything
Isolating business operations from each other is a good instinct, until every operation carries its own client for the same third-party API and its own reading of the same data. Share contracts and integration clients; keep behaviour isolated.
Designing for Human Latency: Task-Based Patient Interactions
In chronic care, the normal latency of an interaction is days or weeks. Model those interactions as tasks, not synchronous flows, and you can control demand, set honest expectations and survive a cohort-sized spike.
Move the Smaller Population When Splitting a Stripe Account
When two businesses sharing one Stripe account need to separate, the question isn't how to move customers. It's which customers to move. Move the smaller, quieter population and leave the growing one alone.
Decommissioning a Telehealth Platform Without Losing the Records
Switching off a platform is easy; keeping its patient records accessible for regulators, lawyers and GDPR requests for another decade is not. How we chose a single final resting place and built a simple, repeatable export.
Carving a Shared Service Out Into a Regional Environment
When a regional business is divested, the shared services it depends on have to move too. A practical checklist for relocating a multi-tenant service and its data: stand up, copy, prune, re-point, verify.
Take Payment as Close to Shipment as Possible
A reusable request → fulfil pattern for products and services, with payment treated as just another eligibility gate. It works the same for direct-to-consumer and employer-funded customers.
2022
Know Thy Runtime
How a four-line change, AES decryption and Ruby's Global VM Lock turned into a P1 incident, and what it taught us about understanding the runtime beneath our code.
This Is How I Get Through a Day at Babylon
Time-blocking, batching context switches, and a modal, terminal-first setup on an iPad Pro: how a Principal Engineer protects focus while bouncing between teams.
2014
Coding In the Cloud [practical fact or developer fiction] part 2
Six weeks of trying to code from an iPhone 6 Plus with a monitor and Bluetooth keyboard: what worked, where the device fell short, and why I moved on to a cellular iPad Air 2.
Coding in the Cloud [practical fact or developer fiction]
Why I moved my primary coding environment to a cloud VPS in 2014: Vim, tmux and mosh over SSH, Dropbox as a safety net, and what it meant for working on poor connections at customer sites.
2013
USP for FSWire Ltd ?
What FSWire offered financial firms in 2013: a multi-stage funnel filter that turns hundreds of millions of daily tweets into a structured, relevant stream, and the case for buying that capability rather than building it.
Big Data - what was the question ?
Reflections on analysing hundreds of millions of tweets a day for financial insight, and why a hypothesis-first approach to big data can't scale when there aren't enough data scientists to ask the questions.
Real-time sentiment & analytics social finance API
Announcing a 1.0 beta API that exposed FSWire's real-time social analytics (sentiment, volume and trending for US and UK securities and FX pairs) as JSON, free for personal use.
To Buy to Sell - this is the Question
Why predicting stock prices from social sentiment alone was naive, and how FSWire moved from raw sentiment to curating context-sensitive content, event detection and a Social Strength Indicator.
URI Rules the World
Why the links inside social posts turned out to be rich signals: expanding shortened URIs, who shared them and when, and how that fed junk filtering, relevance and authority models for financial content.
FSWire Architecture
How the FSWire stack was put together in 2013: Rails on Heroku for the dashboard, and an analytics backend built on RabbitMQ, Redis, Postgres, MongoDB and MapReduce, consuming around 1 TB of Twitter data a day.