Everything, in detail

An operating system for running the business, built on a Workflows–Agents–Tools architecture. Twelve department agents: orchestrator, sales, marketing, CRM, analytics, customer success, web development, market intelligence.

Each agent reads a written SOP, calls a deterministic Python tool to execute, and escalates when the situation isn't in the SOP. The judgment is the agent's. The execution never is.

It runs my own pipeline. 104 prospects moving through a stage-based funnel, over a database layer I can swap without touching a single business tool.

Before an application goes out, a script checks every figure on the CV against one master file. Not a spellcheck. A build step.

That's the thing about a fabricated number. It doesn't feel like lying when you write it. It feels like remembering.

The system around it handles LinkedIn posting ingestion, CV tailoring, ATS validation and Google Sheets logging. It costs nothing to run.

I built the first version on a language model and threw it out. Two identical inputs gave two different recommendations, and I couldn't explain either.

The engine is deterministic now. No model call, no network dependency. The same input produces the same recommendation, and I can walk anyone through exactly how it got there.

Use a model where you need judgment. Never where you need the same answer twice.

Competitor Radar, about 4 hours a month. Every Monday it checks three competitors' live ads, flags what's new, and summarises the angle they're testing.

Reporting, about 6 hours. It pulls the analytics, computes week-over-week deltas, and writes the 'what changed and why' summary a human would.

SEO Engine, about 5 hours. Feed it a keyword and it returns search intent, the questions ranking pages answer, an outline, and internal-link targets.

Lead Scoring, about 5 hours. A new lead arrives, the agent enriches the company data, matches it to the ICP, and scores it 1 to 5 with a one-line reason.

Inbox, about 3 hours. It drafts first replies to DMs and form fills in my voice, with the next step already suggested.

Not 'AI writes my posts'. 'AI runs the process, I make the decisions'. That's the line between a tool and a system.

Research and inspiration. Content strategy. Idea generation. Writing and carousels. Review and optimisation. Publishing and engagement. The analytics loop. An automation layer over all of it.

The blank page was never the hard part. Deciding what goes on it was.

A high-converting site with online booking, so patients book in thirty seconds at midnight without a phone call.

WhatsApp automation: a confirmation, then a reminder before the appointment, on the channel people actually open.

A reputation engine. After each visit a satisfied patient gets a friendly message asking for a Google review.

The loop closes on itself. Reviews lift the Maps ranking, more patients find the practice, the site books them, WhatsApp makes sure they show up, happy patients leave more reviews.

Note on numbers: the percentages usually quoted for message open rates and no-show reduction are healthcare-marketing industry benchmarks, shown as directional ranges. They are not my own measured results.

I ran the editorial calendars for all of Webi Studio's client accounts, publishing 3 to 5 pieces per week per account on Instagram and Facebook. Average engagement rose 35% over three months. Average response time to the communities dropped under two hours.

It was the number we could measure. That's a different thing from the number that mattered.

What I'd measure now: reply latency, and how many conversations turn into an enquiry. Both are harder to pull. Both sit much closer to money.