That pattern runs through everything here: school governance, AI operations, infrastructure, health, sport, games and making. There are 49 entries: 13 full case studies and short notes on the problem, what I built and what it does.
Where it matters, each entry also says whether a model runs when the system operates, or whether AI only helped me build it.
The system that runs SGAF day to day. One agent I talk to, single-purpose specialists behind it, adversarial checks between them, and a decision board that shows me one choice at a time. The mechanisms matter more than the headcount.
Published school data is spread across a dozen releases, each with its own format and caveats. I built one free site that sets every state-funded school in England beside its local authority and England, rebuilt and checked every week, on a map that works from the whole country down to a street. AI agents helped build it. No model runs when it is used.
Every part of school improvement has a system except governance. I designed one: a quality model that says where a board stands, an assurance model that says how it earns the next stage, and a self-assessment that turns both into work a board can do.
A Salesforce-style CRM built inside SGAF: accounts, contacts, a deal pipeline, an append-only timeline, tasks and a nightly health score, with every quote, agreement, invoice and payment captured as it happens and free-tool use tied to recorded consent.
How AI is kept trustworthy inside a product schools rely on: the right model for each job, outputs that are tested or reviewed before anyone relies on them, an impact assessment for the AI processing, and suppliers vetted against written terms.
Dense documents arrive at a board and nothing happens. These systems turn an inspection-data PDF, an inspection report and a school improvement plan into reviewable evidence, source-linked questions, released governor papers and delegated work that comes back.
Every change goes through a pull request with build, lint, type and test gates in a fixed order, a check that no work is stranded, and database migrations confirmed before merge. The merge is the deploy, so the gate is the release process.
A real courtside problem became a native mobile app: prepare the match, run the clock, score live, track the centre pass, correct deliberately, and share a clean result at the whistle. AI helped build it. No model runs when it is used.
A mini PC running Omarchy became the build, database and test machine for SGAF: the whole stack on one box, synthetic data only, AI agents in separate branches, and a move from the Mac planned so that nothing is lost.
Phone and wearable health data became normalised records, a daily coaching plan, prompts on my phone, an adherence log and a weekly review that corrected the coaching itself. The loop is the point, not the chatbot.
The public funding finder was rebuilt from official sources across the four UK nations, with every record carrying its source, check date and availability. A maintenance agent was built to discover, recheck and expire programmes, with a person approving every change.
I use AI to adapt 3D models, then print, inspect and revise them. A pegboard mount for the mini PC that runs my development server, printed on a calibrated profile.
Before the current front-door operation, I built and tested an agent organisation on an open-source agent framework: one Chief of Staff routing durable tasks to six specialists, with an independent Assurance role returning an auditable verdict.
Each app solves one governance problem end to end and feeds the next. Together they are one assurance system for school boards and academy trusts.
A free self-assessment that marks every part of governance Secure or Developing, captures timestamped evidence against each element, and produces a report for board papers. Almost every other SGAF app feeds it or reads from it.
No model at runtime
Upload the school improvement plan and it becomes a year of structured governor visits: one named governor per priority, three phases, Word visit packs, a year at a glance, and visit reports that governors complete by private link and reviewers approve. Statutory monitoring roles run on the same engine.
Model in parts, checks deterministic
Agreed governance actions in one place. Each task has success criteria, an owner and a deadline, reaches the governor by private link, and is only completed once the school has reviewed the response against the criteria.
No model at runtime
Short governance checks across safeguarding, finance, attendance, SEND, cyber security, data protection and more. Each gives a rated posture, a check can go to the colleague who knows the area by private link, and weak results become named risks or planned visits.
No model at runtime
The board's single log of what could stop the school meeting its aims, with scores, owners, controls and review dates. Weak audits and worrying data trends suggest risks straight into it, and across a trust it rolls up into one heatmap.
No model at runtime
Every area the board must assure gets a review, a term and a lead governor. Review Cycle holds the whole three-year programme, tracks each review from planned to closed, and lets the board decide what reaches it.
No model at runtime
The board's meeting rhythm across the year, with each agenda pre-loaded with its standing items and the statutory items due that term, suggestions drawn from the school's own intelligence, and a branded Word agenda with a matching minutes template.
Reads a board's minutes and scores the challenge they record across the dimensions of strategic scrutiny, with an outside-eye view that never quotes a name or a line. Personal details are redacted before scoring.
Uses a model at runtime
A structured termly report in eleven steps, shaped to the school's phase, that puts the school's internal picture on record for the board and feeds the Board Intelligence Report. Improvement priorities pull through from the uploaded plan.
A termly board paper that combines GIAS, Ofsted, Companies House, DfE benchmarking and performance data with the school's own records, reviewed before it is saved as a dated version. A section with no current source says so instead of guessing.
Builds the board's annual governance statement from the year's approved visit reports, completed checks and Headteacher Reports. The board adds its own perspective, agrees the wording and records adoption, which freezes the version it approved.
Attendance, behaviour, outcomes, Pupil Premium, SEND, staffing and finance beside the local authority and England across several years, with a plain explanation of every measure and a meeting brief of questions with owners and review dates.
A school's published results beside England and the nearby median for the same year and phase, with a nearby set or a wider set matched on phase, roll size and free school meals, and every definition and group size in plain sight.
No model at runtime
A school's recorded SEN beside the local authority and England, or nearby schools within a chosen radius, turned into neutral context and ten questions the board can ask leaders, ready to copy into board papers.
A termly scan of a school's website pages and documents that records what it detected, what it could not resolve and what needs a person, then keeps a prioritised issue queue with evidence, delegation emails and a history across the year.
Model in parts, checks deterministic
Lists the policies a school is expected to hold or publish, matched against the latest website scan, with obligation, status, owner and next review. The scan suggests; a person confirms whether each policy is public, held internally or not held.
Holds an academy trust's central website to the same standard as its schools' sites: a scan against trust-level publishing duties, then a person confirms each duty with where it is published, kept term by term.
One view for trust administrators before the trust board meets: a school-by-school briefing, the Quality Standard across every school, Headteacher Reports, visits, website checks, review cycles and risks, with exports for board papers and a term handover that keeps its reasoning.
No model at runtime
Portal-wide help that combines reviewed library content with page context, short steps, verified links and support drafts. Answering from a school's private records is deliberately switched off, and live use drove the next round of improvements.
Uses a model at runtime
A tool that extracts priorities and questions from an inspection report while keeping report dates and source links, and handles predecessor-school evidence separately so an old finding cannot look current.
Model in parts, checks deterministic
Plan extraction feeds operator review and an explicit release step. Governors get the reviewed plan while a later revision stays separate. Each priority becomes a page with milestones, leads and questions.
Model in parts, checks deterministic
An uploaded inspection-data report becomes reviewable measures with page provenance and short views for different board audiences. One example extracted a hundred and sixteen measures and chose thirteen for the short view.
Model in parts, checks deterministic
What runs underneath the apps: the CRM, official data pipelines, release gates, monitoring and data protection.
Each client meeting gets its own HTML deck, served from the SGAF portal itself, so a demonstration moves from the argument to the live product and back without changing windows. Keyboard and clicker control, full screen, a link to every slide and print-ready leave-behinds.
When something in the product goes wrong for a school, support should know before the school has to write in. A small set of product signals prompt a follow-up. Repeats are suppressed and every follow-up is a draft.
Model in parts, checks deterministic
Public education and community datasets arrive on their own schedules. The pipeline checks versions and hashes, validates candidate releases, keeps the last good one and records freshness receipts, so a failed import never becomes trusted product data.
No model at runtime
Official guidance changes are connected to the lessons and resources that depend on them. The process separates a substantive change from page churn, and a review date from an actual modification.
Model in parts, checks deterministic
Schools stay in control of their data: privacy notices that match practice, a data processing agreement with named sub-processors and a right to object to changes, AI processing that never trains on customer data, and an impact assessment for the AI features.
A layer that detects missing publications, failed or retried jobs, broken evidence handoffs and test records that could contaminate the operational picture, with bounded heartbeats and read-only reconciliation.
No model at runtime
How the agent organisation is designed, what it coordinates, and what it is not allowed to do.
Products with a real job and no model at runtime.
Content as a system: generated catalogues, gated drafts, and resources that stay current.
A media library, school-calendar themes, a long posting plan and the product facts in one place. Approval fingerprints the exact text; the posting agent acts only on an approved match and reports the live link back.
Model in parts, checks deterministic
Brief, grounding, authoritative sources, draft, render, independent review, final edit. Brand rules are checks that fail, and the author cannot approve its own output.
Uses a model at runtime
One catalogue generated from the real articles, resources and films, shared by training, help and the posting plan. Help content lives beside the product, and missing media or a rendering failure can fail the build.
No model at runtime
Six hundred and fifty-two school-work prompts with stable identifiers, a page each, a school profile that fills them in, and a record flag on anything that could become a pupil, staff or legal record. A person still owns the draft.
No model at runtime
Markdown posts, generated social cards, a full-text RSS feed, cover images and copy controls that turn any post into an article for X or LinkedIn. The site is built and previewed on my own server before anything is pushed.
No model at runtime
Machines and parts that do real jobs: a build server, a site watchdog, and parts designed with AI and printed.