Canberra · Australia
I build the AI systems a business actually keeps — and I make sure it still runs after I've gone.
Rather than describe it
An enquiry arrives. The agent reads it, classifies it, routes it and drafts the reply — then stops and waits for a person. You're the person.
01 — Who I am
Spain, Nepal, California, Australia. Four countries, four trades, one habit: I go and look at how the work is actually done before I change it. Open any year.
Turning what I built inside the clinic group into modular automation other multi-site operators can deploy — dental, physio, psychology, GP, allied health.
An item-code optimiser that recovers $5–15K a month in missed billing. A content engine that runs unattended. Reporting that builds itself.
We moved to Canberra when my partner took a clinical management role here, and I joined the same group on the operations side.
Twenty-plus staff and every process held together by spreadsheets and memory. I started automating pieces of it to survive my own workload, and ended up designing and building the entire operations and AI stack — on my own, while running the day-to-day.
I'm not there any more, and all of it is still running.
My partner suggested Australia — we had to enter before turning thirty to qualify, so we went. A year in Sydney: sales director at a carbon consultancy, then sales and marketing for an events business.
It taught me to explain a technical idea to someone with no interest in the technology and every interest in the result. That turns out to be most of the job when you're the person introducing AI into a business.
My partner and I started a small botanical care business in Valencia and ran it for a year.
Alongside it I took on the family property portfolio — and later built its software myself: a bilingual owner portal where owners see their properties and financials directly, plus an intake agent that reads documents and books maintenance jobs and owners' meetings. Still running today.
COVID made me want a change, and an American developer took me on. On site I supervised and coordinated the renovation works. Off site I coordinated the team still working from Valencia — marketing, the rentals side, the designer — across a nine-hour time difference.
It's where I learned to run a project without being in the room with the people doing it.
It ended over visas. Continuing would have meant staying on the wrong side of the rules, and I wasn't willing to do that, so I came home.
Comparative market analyses. Winning and maintaining listings. Qualifying buyers against real financial capacity. Contracts and paperwork. Campaigns.
Three years of doing, by hand, almost every task an agency later asks someone to automate. I learned the part of real estate no drawing teaches you: it's a business of a hundred small, repetitive, deadline-driven jobs, and the people doing them are almost always underwater.
I did my final university project in Gahare, Nepal — not on paper. Five months coordinating phase two of a primary school rebuilt after the earthquake: two classroom buildings on split levels, retaining walls, landscaping, leading volunteers from six countries. Traditional bioconstruction combined with seismic-resistant methods.
No budget padding, no consultants, no safety net. It's still the most complete design-to-build project I've been part of, and it's where I learned what a system costs when it fails.
Back home I co-founded a small practice — a town hall building, a penthouse renovation, a residential development. Concept to site.
How to hold an entire system in your head, and how a change in one place moves through everything else. Structure, sequence, constraint, tolerance, dependency.
The vocabulary is identical when you're designing an automation layer instead of a building.
Everything I built there is still running. I don't work there any more.
02 — What I've built
Ten systems I designed, deployed and handed over. Different industry to yours, probably — but the shape of the problem rarely changes. Open any entry.
Three people used to sort a shared inbox every morning. The agent classifies each message by location and intent, routes it, and drafts a reply where one is needed.
It never sends on its own. A person approves — and the system records whether they agreed or overruled it, so it improves and accountability stays human.
n8n · Gemini · Gmail API · structured output parsing · approval step
The platform at the centre of the business had no usable API — by design, because the vendor sells the lock-in. I didn't fight it and I didn't replace it. I built around it: scheduled structured exports, a small extraction service, a normalisation layer.
It ran every day for two years, and every dashboard and agent was fed by it.
Headless extraction · scheduled exports · ETL service · Docker
Timesheet ingestion, healthcare award interpretation, exception flagging, payroll-ready output across six locations.
The forty-five minutes left are deliberate: a person reviews the exceptions with everything already assembled. Automating that last mile would be reckless, not clever.
n8n · MYOB integration · award rule engine · exception queue
Built for the portfolio I still look after. It takes documents in by email or chat, reads multi-page files, works out what they are and files them — and where the content implies an action, it creates the calendar entry.
Nothing is committed without a person confirming. Absorb the routine volume, surface the decision, keep someone accountable for the call.
n8n · multi-page document reading · chat + email intake · calendar actions
Owners log in and see their properties, documents and financial position directly, instead of waiting for someone to assemble a monthly report. Data model, authentication, interface and hosting, all mine.
Small on purpose. The point is that owner reporting stops being a recurring manual task and becomes a live surface.
Custom web app · authentication · bilingual ES/EN · Docker · self-hosted
Reads completed treatment records against the item code schedule and flags what was delivered but never billed, or billed under the wrong code.
No new sales, no price increase. Recovered margin on work already done.
Structured extraction · rules + LLM validation · exception reporting
Chair utilisation, recall rate, revenue per patient, treatment acceptance, new patient count — per site and consolidated. Nobody compiling spreadsheets on a Friday.
Leadership stopped asking for reports and started asking better questions. That was the actual return.
ETL pipeline · Supabase · dashboard layer · automated distribution
Supplier invoices arrive, get matched against expected purchases, checked for duplicates and approved automatically below a set threshold. Anything unusual goes to a person with the reason attached.
Document parsing · supplier matching · duplicate detection · approval routing
Ideas generated from a configurable brief, drafted in a defined voice, pushed to a phone for a yes or no, then scheduled and published.
The pattern generalises straight to property copy, suburb market updates and campaign variants.
n8n · LLM generation · Telegram Bot API · Sheets config · scheduler
Containerised services behind a reverse proxy with automatic TLS, on a server I administer. No per-seat tax on the automation layer, no vendor holding the data.
Maintained like production, because it is: every workflow backed up to version control nightly, and a daily report listing any failure with a direct link. Automation nobody is watching is just technical debt with better marketing.
Docker · Traefik · Linux · self-hosted n8n · Supabase · nightly Git backups
03 — Where I am now
Three things running right now. Every one of them started life as somebody's manual process.
Turning the stack I built inside a six-clinic group into modular systems other multi-site operators can deploy — dental, physio, psychology, GP, allied health.
The item-code optimiser alone recovers five to fifteen thousand a month in treatment that was delivered and never billed. No new patients, no price rise.
dentalbee.com.au ↗Several companies at a time, a few hours a week each. Most of it sits under confidentiality agreements, so it will never appear on this page.
The pattern doesn't change much: find the process everyone quietly hates, prove it can come off their desk, then build the thing properly.
Live return-per-project across a dozen concurrent jobs, where each one runs six to nine months and the costs arrive in no useful order.
Slow-burning data — exactly the kind nobody can hold in their head, and exactly the kind that changes which work you take on next year.
I don't buy another platform. I build the piece that's missing.
Which means it belongs to you — your server, your data, your credentials. Nothing switches off because a subscription lapses, or because I stop answering the phone.
04 — What this looks like in your business
I don't know your industry yet. But I know what this work looks like when it's done by hand — because I did it by hand for years, in three of them.
Enquiries arrive by email, web form, phone and social, and someone spends their morning sorting them. An agent classifies each one by intent and urgency, routes it to the right person and drafts the reply. Nothing sends until a human presses yes.
Replaces — three people triaging a shared inbox before 9amQuotes, invoices, contracts, timesheets, reports — arriving as PDFs and phone photos, then getting retyped into a system by a person. Read them once, pull the fields, file them, and surface only what doesn't match.
Replaces — retyping the same numbers into a second systemLive dashboards fed automatically from the systems you already run, per site, per team, per project and consolidated — instead of somebody assembling a spreadsheet on a Friday afternoon.
Replaces — the weekly report that arrives too late to act onWork delivered and not invoiced. Invoiced under the wrong code. Discounted by accident and never noticed. Reconciling what was done against what was charged finds margin without a single new sale.
Recovers — revenue on work you have already doneIdeas in, drafts out in your own voice, approved from your phone in ten seconds, then scheduled and published. Measured afterwards, so next month's is better than a guess.
Replaces — the content that never quite gets writtenMost software a real business runs on was never built to be connected — no API, or one that costs more than the problem. I get in the other way: scheduled exports, document parsing, or teaching the AI to drive the screen the way a person does.
Unlocks — the platform your vendor won't openAny business where the work is repetitive, deadline-driven and currently held together by people remembering things. I've built inside multi-site healthcare, real estate, property management, architecture and construction, events and professional services.
Best somewhere between ten and a hundred people, with more than one location or more than one system. It works badly when nobody inside the business actually wants it — and I'd rather say that in the first conversation than in the third month.
You describe the process that hurts most. I tell you honestly whether it's a two-week build, a two-quarter one, or something you shouldn't automate at all. No charge and no slide deck.
I work alongside your team, design the system and build the pieces that need building. This is how I work with most businesses today.
Owning the whole automation layer end to end. What I did for two years at a six-site group — and everything I built there is still running without me.
How I'd do it
Most automation projects don't fail on the technology. They fail because nobody watched the work first, or because nobody owned it afterwards. None of these six are about AI.
A week beside the people doing the work — not a workshop. What they describe and what they do are never the same, and the gap between them is where the automation goes.
Small, and visible to whoever feels the pain. Nobody should take automation on faith for a quarter. The first win buys the trust that pays for the harder work behind it.
Highest volume, clearest rules, fewest people to convince. Lowest risk — and it produces the clean data the next system needs anyway.
The question is never whether a platform has an API. It's what a person can do inside it, and how they get data out today. A report, an export, a screen — that's the door.
Handle the repetitive band completely, then put a person on the edge cases with the context already assembled. Speed where it's safe, judgement where it isn't.
Infrastructure, data and credentials belong to the business. Backed up nightly, failures reported daily. Automation nobody is watching is just technical debt with better marketing.
Toolkit & credentials
05 — The cost of waiting
That's what makes this decision hard. The cost never arrives as an event you can point at — it arrives as a slope. Five things that are already happening if you haven't started.
You may have seen the aftermath: a provider who promised a lot, invoiced, and left behind something nobody uses. It happens for the same three reasons every time — bought before it was understood, built by someone who then left, owned by nobody inside the business.
The point was never to have AI. The point is that in a year the thing is still running — and that it belongs to you, not to me.
Tell me the one that hurts most. I'll tell you honestly whether it's a two-week build, a two-quarter one, or something you shouldn't automate at all.
Contact
Based in Canberra, working with businesses anywhere. The most useful first message is one paragraph about the process that hurts most — no brief required.