Projects
Each entry records the problem, the decisions that mattered, and the measurable outcome. Filter by discipline below.
Led end-to-end development of review-response agents on OpenAI and Claude, with tool integrations and memory for autonomous operation — and a prompt framework that cut model cost 18%.
Resilient integrations with Google Business Profile, Yelp and Facebook Reviews — webhook-driven where the platform allows it, scheduled where it doesn't, reconciled either way.
A multi-locale marketing site moved onto a Next.js frontend with WordPress as the editing layer — including the preview and publishing pieces that headless builds usually drop.
React and Next.js dashboards surfacing AI metrics and generated responses for client decision-making — increasing engagement 28% by making the model's work inspectable rather than opaque.
A RAG layer over each business's own details, so generated responses referenced things that were actually true — and said nothing when the source material didn't cover it.
The interface where humans approve, edit or reject AI-generated responses before they go out under a client's name — built so the checking is faster than writing the reply yourself.
Dockerised CI/CD onto Kubernetes for cloud-native Python and Node services, plus a data-model redesign that doubled pipeline throughput for AI processing.
A pipeline turning raw review text from several platforms into reputation insights — and the calibration work needed before anyone would act on the numbers.
Rebuilding a bloated theme around a constrained set of sections — so campaigns ship without a developer, and the storefront stops being slow because of apps nobody uses.
Turning years of free-form HTML into validated structured fields, without losing a single URL — and rehearsing the whole migration against production data before cutover.
Instrumenting a multi-service SaaS platform so problems announced themselves instead of being reported by customers — and the incident process that turned alerts into fixes.
Replacing per-service authentication with one issued-token model — so that adding a service no longer meant reimplementing authorisation and hoping it matched.
Internal system-health and analytics screens over a multi-service SaaS platform — built so an on-call engineer could tell in one glance which service was at fault.