Case studies
Twelve systems, each told the same way: the problem, what I built, and what changed. The work studies are sanitised (no client names, no internals); the personal ones are fully mine, so they can be shown in detail.
At work · a hospitality feedback analytics SaaS company
A The one number that kept the biggest client An industry-first composite metric, born from years of sitting with the clients who needed it, that fused AI sentiment analysis with operational ratings into a single score operators could run a business on. retained the client worth 50%+ of revenue
B An AI teammate that works the ticket backlog A three-skill workflow that specs, decomposes, and orchestrates a team's Jira backlog, letting one developer run several epics in parallel while agents work the well-specified tickets. one epic per fortnight became four epics in parallel
C An AI that tests its own work in a live cloud environment A tester agent that syncs finished work into an isolated cloud environment, executes it in the real runtime, and reacts only to structured, machine-readable results. edit-to-verified loop with no human in the middle
D Replacing a full-time reporting role for 250 branches When the biggest client made replacing a manual reporting job a condition of renewal, a paginated-report engine took over the work, running seven days a week instead of five. a £30,000+ a year manual process, automated and expanded
E AI-driven reliability engineering for an overnight data platform Months of run history audited by agents into a defensible failure-mode taxonomy, then a self-observation layer so the platform reports its own nights. from "we find out when a client complains" to a daily evidence-based digest
F The reporting platform: 400 branded reports without hands A nightly notebook engine that generates, quality-checks, and delivers hundreds of branded client reports, plus a generative skill that builds new Power BI reports from a one-line prompt. 30 to 80 hours of weekly manual work replaced by a run that takes minutes
G Fleet operations: 170+ semantic models, refresh, rebinding, and cost control One overloaded all-client model became 171 per-client models maintained from a single baseline, ending daily capacity upgrades and putting cloud spend under active management. near-daily capacity upgrades stopped the week it shipped
On my own time · personal projects and infrastructure
H Courting: a three-platform product shipped solo, with agents doing the shipping A racquet sports management platform on iOS, Android, and web, delivered through a 14-skill agentic pipeline that specifies, implements, tests, and releases. live product; 2,100+ tickets and 2,000+ PRs with agents in the loop
I graphify: giving agents a knowledge graph instead of a file search A self-built CLI and MCP server that turns code and documents into a persistent, queryable knowledge graph, wired into the agent harness. structure once, query forever
J The adversarial plan review, generalised and ready to open-source The cross-model review pattern rebuilt as project-agnostic, reusable skills, battle-tested on personal projects. the lead candidate for public, inspectable proof
K A memory and governance layer that makes the whole agent fleet smarter Persistent atomic memory plus written precedence rules, so agents start every session knowing the environment and humans only approve what matters. sessions start already knowing the sharp edges
L MindGrid: a consumer mobile game, AI-assisted, off the Microsoft stack A React Native word-puzzle game with a Supabase backend, a tiered test harness, and its own internal level-generator tool. shipped consumer software on a second stack