I turn operational problems into working AI tools and digital products.

A small therapy practice runs on a surprising number of disconnected parts: booking, intake, consent, session notes, treatment plans, homework sent to clients, invoicing, subscription limits, marketing pages and email follow-up.

Each of those has a tool. None of them share a client record. The therapist becomes the integration layer, re-entering the same details and reconciling the same appointment across four systems while trying to deliver clinical work.

The problem was never a missing feature. It was that no single system held the client, the plan, the sessions, the money and the communications together.

Building it as one product with a shared data foundation meant the operational workflows could finally reference each other: capacity affects booking, booking affects billing, the treatment plan drives the session workspace, and the session drives what the client sees in their portal.

Marketing performance lived in Google Ads, GA4 and Search Console. Commercial reality lived in a CRM that had been inherited and never fitted the business. Website enquiries arrived by email and were retyped by hand.

The question everybody wanted answered — which marketing activity produces real opportunities — required a person to manually reconcile four sources, so in practice it went unanswered.

The gap was not analytics. Each individual tool reported accurately. The gap was that no system joined a keyword or a campaign to a deal with a commercial value attached.

Once enquiry ingestion, deduplication and source tagging were automatic, attribution stopped being a reporting exercise and became a property of the record itself.

Monitoring rankings at any scale means paying per query to an external data provider on a schedule, for many keywords, across many customers.

That turns an apparently simple product into a cost-control problem: without frequency awareness and usage limits, the margin disappears quietly and only shows up on the invoice.

The interesting engineering is not the chart. It is the scheduling, the tenancy and the spend controls that make the chart affordable to produce.

Building the operator console alongside the customer product meant the commercial model was visible during development rather than discovered afterwards.

Selling through a website, a shop counter and a team of sales managers means three different order intakes, three views of the customer and one accounting process expected to absorb all of it.

Loyalty and repeat purchasing are where the margin lives in this category, but they are impossible to run well when the customer record depends on which channel the person happened to buy through.

The accounting workload was a symptom rather than the problem. The real cost was that no single place described a customer or an order across the channels.

That distinction is why the work splits in two: automating what already existed and could be measured, and prototyping the unified system separately rather than pretending it was built.