Applications and automation, with the evidence
I build and run a portfolio of applications and automations end to end: SGAF and its tools, the AI agents that operate them, and the data and release processes underneath. Every line below links to work you can open.
- Pull requests merged
- 962
- Automated test files
- 1,348
- Database migrations
- 249
- Written decision records
- 379
Counted from the SGAF repository on 18 October 2026, since its first commit on 1 March 2026.
- Schools have used SGAF's free tools
- 500+
- Free checks completed
- 1,200+
- Schools have run the School Data Check
- 340+
SGAF in use: completed free checks from 3 April to 18 October 2026, counted by school and rounded down.
Application portfolio and roadmap
Knowing what the estate is for, cutting what duplicates, and building only what earns its place.
- Restructured a team of forty-five AI agentsRemoved the overlap and gave every remaining role one written job, an owner and a list of things it must not do.
- SGAF, app by appEach app solves one governance problem end to end and feeds the next, from the self-assessment to the board paper.
- A bespoke CRM built on the product's own dataAccounts, contacts, a pipeline, a timeline, tasks and a nightly health score, built additively so nothing was lost.
- One generated content catalogueReplaced three conflicting lists. Missing media now fails the build instead of going unnoticed.
- Built where the facts already livedPut the posting plan inside the CRM rather than buying a scheduler that could not see the product facts.
Automation and integration
Automation that is prioritised, governed and connected to the data it depends on.
- One backlog for competing agentsCapturing an idea is separate from authorising work, and nothing is built until it is promoted.
- Official data, integrated safelyVersions and file hashes checked, each release validated, and the last good copy kept.
- A year of visits from the school's own planThe improvement plan becomes a visit programme: one governor per priority, packs, reports by private link and approval.
- Event-driven product signalsProduct events reach support without ever blocking the product, with repeats suppressed.
- Documents into structured evidenceInspection data, inspection reports and improvement plans turned into reviewable work for governors.
- Official sources across four nationsA funding catalogue where every record carries its source, its check date and its availability.
Service, resilience and release
Keeping applications available, current and safe to change.
- Release gates on every changeBuild, lint, type and test gates in a fixed order, migrations confirmed before merge, and the merge is the deploy.
- Monitoring the automation itselfDetects missing publications, failed or retried jobs and broken handoffs, and reports without repairing.
- Uptime monitoringOutages and recoveries reach my phone through the same workflow as everything else.
- A dedicated build and test serverThe whole stack on one Linux machine, synthetic data only, and a migration that moves nothing until it is verified.
- AI help, evaluated after releaseJudged on real use rather than the demonstration, which showed exactly where routing and follow-up needed work.
Governance, risk and suppliers
Clear decision rights, checks that can fail, and suppliers held to written terms.
- Adversarial gatesEvery producer has a checker with a reason to disagree, and the last gate is always a person.
- Decision rights stay humanAgents prepare and draft. I decide, send, spend and delete.
- A governance quality frameworkCompliance, Assurance and Continuity, with an assurance model that holds every priority to the same standard.
- Risks suggested by the board's own workWeak audits and data trends arrive as draft risks the board accepts or dismisses, and roll up across a trust.
- AI governance in SGAFThe right model for each job, outputs tested or reviewed before anyone relies on them, an impact assessment and vetted suppliers.
- Data protection and sub-processorsNamed sub-processors, notice of any change, and AI processing that never trains on customer data.
- Governing AI the way a board governsFour questions for any work an agent touched: who set the limit, what is the proof, where is the record, who signed.
Data, SQL and reporting
Turning published data into something a board can question.
- One board paper from many sourcesGIAS, Ofsted, Companies House and DfE data with the school's own record, reviewed and saved as a dated version.
- SEND data for governorsSchool, local authority and national data with nearby comparison, needs breakdowns and trends.
- School comparison with contextOutcomes beside pupil and area context, with sources and dates visible and no misleading league table.
- Inspection data into briefingsOne report became 116 measures with page references, with 13 chosen for the short view.
- Policy changes mapped to resourcesOne review mapped 33 official sources to 34 resources and produced nine substantive updates.
Leadership and stakeholders
Leading people, setting direction and explaining technology in plain terms.
- More than twenty years in educationIncluding a school senior leadership team, so I know how the people using the systems actually work.
- A demonstration to an established edtech companyWalked its team through a running autonomous operation, including the honest scorecard.
- A bespoke deck for every client meetingHTML presentations served from the product itself, so a demonstration moves between the argument and the live portal.
- Roles, objectives and review for an agent teamWritten jobs, clear ownership and a decision board that brings me one choice at a time.
- Teaching that stands up to scrutinyTraining resources checked against Ofsted and Department for Education sources before they publish.
What it is built with
TypeScript, Next.js, PostgreSQL and SQL, APIs and webhooks, scheduled automation, AI models with evaluation, and Linux. The stack page lists the rest, area by area.