My products

Two open-source tools for the part of the workflow nobody demos.

Both are published, both are permissively licensed, and both ship with the benchmark harness so the performance claims can be re-run rather than taken on trust.

CI Kit

Run your CI on your laptop, before you push.

v1 released — MIT

CI Kit gives a repository a local gate that mirrors what the GitHub Actions workflows do: the same installs, the same linters, the same type checks, the same tests. It also runs the workflow files themselves locally through act, so the YAML can be checked without waiting for a runner. The design principle is a single copy of the setup: the gate scripts live in the repository under scripts/ci/, and both the Makefile and the workflow call them. There is one copy of the setup, so it cannot disagree with itself. The tool works with any language — the gates are driven by what the repository actually contains — and it deliberately does not take over your pipeline. The workflows stay the source of truth; the kit makes them cheap to try.

99 son a 2022 Intel laptop, against 43–48 s on a runner
43–48 sGitHub runner, measured on identical work
28 of 30runs that failed without ever starting a job
3in under four minutes, zero runner minutes
One copy of the setupRuns the workflow locally via actLanguage-agnostic gates
  • A feedback guarantee that a billing decision cannot switch off

    On 21 October 2026 the largest repository in the workspace had 28 failing runs out of the last 30, none of which failed on code — GitHub had refused to start the jobs because of a spending limit. Every run failed in about six seconds. The local gate answers the same question in about a hundred seconds, and it says why.

  • Three real defects found in under four minutes, without a runner

    A stale secret-scan allowlist entry that was failing the CI scan on a tracked documentation file, ninety-three dependency advisories of which eight were critical, and build artefacts tripping a working-tree secret scan. The first two were masked because CI was red for a completely different reason, and nobody could tell the two apart.

  • The claim is deliberately narrower than the marketing would be

    The evidence document has a section titled 'What is not claimed'. It states plainly that the local gate is not faster than a runner, that no runner-minute savings are claimed, and that the numbers come from one machine and one workspace. The value of the tool is a guarantee of feedback, not a speedup — and the documentation says so before anyone asks.

  • CI/CD
  • GitHub Actions
  • Developer Experience
  • Shell
  • Local-First
ShellMIT

api-mock-services also known as SaaS-LocalStack

Stand-ins for the outside world.

v1 released — MIT

Every project that talks to a payment gateway, an email provider, an accounting service, a DNS provider, a messaging service or an identity check needs something to talk to in a test. This is that something: one implementation, shared, instead of a copy in every repository. There are no dependencies. The mocks are built on Node's own HTTP server, so they run in a test, in a container and on a laptop without being installed first. The rules that keep it light are not preferences — a contributor who breaks one of them breaks the promise in the first line of the README — and the footprint is enforced as a build gate, so the promise is checked on every build rather than asserted in a paragraph.

50–51 MBevery mock running, one process
250 MB214 MB of it the OpenAPI proxy
0–0.5%against 7.9% for the four-container stack
0.4–1.4 sagainst 20 s and rising for containers
Zero dependenciesOne process for every mockIn the test, not beside it
  • A fifth of the memory and a fraction of the idle CPU

    Measured on the same machine with every mock running: one process at 50–51 MB and 0–0.5% idle CPU, against four processes at 250 MB and 7.9% idle CPU — almost all of it an OpenAPI proxy that also never settles. The engine that answers the first request in 0.4–1.4 s replaces one that took 20 s and was still rising. The ranges are the real figures: the script that produces them reports a different number on a busier machine, and both documents now say so rather than quoting a sample that happened to be true once.

  • Eighteen providers, including the African payment and identity rails

    Stripe, Paystack, Interswitch, Flutterwave and Monnify for payments. YouVerify, Smile ID and IdentityPass for identity. Twilio, Termii, OneSignal, WhatsApp and FCM for messaging. Resend, Xero, Cloudflare, OpenAI and S3 for everything else. Every mock answers /health, so a container can be checked without knowing which mock it is.

  • The promise is a build gate, not a paragraph

    The footprint test fails the build if the engine exceeds 100 MB, 5 seconds of startup or 1% idle CPU. A promise nobody checks erodes. This one is checked on every build, which is why the numbers on this page can be stated without hedging.

  • Mocking
  • Offline
  • Payments
  • FinTech
  • Testing
  • Node.js
  • Zero-Dependency
JavaScriptMIT

Automation Gateway

Air-gapped workflow execution as declarative JSON, with an audit trail that stays on the host.

In development — repository to be published

An air-gapped, zero-dependency automation platform built to run as a continuous background Windows Service. It decouples orchestration from business logic completely: multi-stage pipelines are declared as data-driven JSON blueprints and executed inside non-blocking background memory sandboxes, while the engine core stays immutable and composable. Nothing in the request path needs the internet, a cloud AI gateway, or a package manager — the platform ships as a folder, is registered as a service, and runs.

0ships as a folder — no package manager
JSONdeclarative blueprints, not controller code
*.plugin.ps1auto-discovered — the core is never edited
Per attemptRetrying, Success and Failed written to SQLite
Data-driven blueprintsPlug-in auto-discoveryTransactional audit trail
  • Business logic became reviewable data

    A pipeline is a JSON blueprint that can be diffed, audited, replayed and approved. Adding a step no longer means a controller change and a deploy — which is the wrong tool for a business rule.

  • The engine core is never edited to add an integration

    Dropping a *.plugin.ps1 file into the plugins directory is the whole extension procedure. The engine core is a separate module scope with a composition root, so a plugin cannot reach into it.

  • Every attempt is recorded, including the retries

    A step that throws is retried with exponential backoff, and each attempt writes a distinct state row. An auditor can trace the transient failure, the recovery delay, and the final outcome without reconstructing anything from logs.

  • PowerShell
  • Windows Service
  • Air-Gapped
  • Orchestration
  • SQLite
  • RBAC
PowerShellMIT

Filter

Filter by technology

Small datasets, real behaviour — the same partial-refresh pattern is reused for the blog topic filter and the evidence metrics.

CI Kit

A local gate that runs the same scripts your GitHub Actions workflow runs, so the answer to 'did I break something?' never depends on a runner being available.

  • 99 son a 2022 Intel laptop, against 43–48 s on a runner
  • 43–48 sGitHub runner, measured on identical work
Repository

api-mock-services

One process, zero dependencies, and eighteen mocks for the payment gateways, identity checks and messaging providers your product talks to — running in the test, not beside it.

  • 50–51 MBevery mock running, one process
  • 250 MB214 MB of it the OpenAPI proxy
Repository

Automation Gateway

A Windows Service orchestration engine that runs enterprise workflows from JSON blueprints — no internet, no cloud dependencies, and a transactional SQLite record of every step attempt.

  • 0ships as a folder — no package manager
  • JSONdeclarative blueprints, not controller code

Principles

The rules I build by

These are not values on a wall. Each one exists because ignoring it produced a measurable problem at some point.

Measure, then claim

A benchmark is a product with a maintenance burden. I publish the harness alongside the result so a reader can re-run it and disagree with me — and I publish the measurement I threw away, because a reader who knows the spread can trust the number.

State what you do not claim

Every evidence document I write has a section listing what is not claimed. A tool that overclaims about speed is a tool that has to be defended, and the defence is always weaker than the original claim would have been.

Configuration is data, not code

A workflow declared as a JSON blueprint can be reviewed, diffed, audited and replayed. A workflow expressed as a controller method can only be changed by a deploy — and a deploy is the wrong tool for a business rule.

Degrade, do not depend

If a component is unreachable, the system has to have a defined fallback and a defined audit record of having used it. A local model with a deterministic fallback beats a cloud model with a timeout.

Background

Experience

  1. 2026 — present

    Open Source Architect

    Independent

    Maintaining CI Kit, api-mock-services and the air-gapped automation engine; writing about CI/CD drift, local-first infrastructure and reproducible measurement.

    • CI Kit
    • api-mock-services
    • Automation
    • Writing
  2. 2024 — 2026

    Platform / DevEx Engineering

    FinTech & SaaS product engineering

    Owned the internal CI platform and developer onboarding for payment-product teams. The workspace these tools came from had 28 red runs in a row and nobody could tell that the jobs had never started — which is what turned a local gate from a convenience into the answer to 'did I break something?'

    • Platform
    • DevEx
    • Payments
  3. 2022 — 2024

    Backend / Infrastructure Engineer

    Product engineering

    Service and infrastructure work across Go, Python and Node services, with a gradual drift toward build tooling, container images and everything that sits between the laptop and the cluster. The Docker and infrastructure-as-code writing from this period is in the earlier archive.

    • Go
    • Docker
    • CI
  4. 2013 — 2022

    Application Support & Software Engineering

    Equipping People (London), Exclusive IT Solutions, Canaries Solutions

    Application support, service desk and delivery work before and alongside engineering — the period in which I learned to read logs, reproduce faults and write the documentation that made a system survivable by the next person. The 2013–2025 archive is an honest record of that arc.

    • Support
    • Delivery
    • Documentation