I built the Michael English Intelligence Brain because I kept watching regulated Irish firms — solicitors, accountants, family businesses turning over fifty or a hundred million — quietly hand their working knowledge to whichever cloud chatbot their staff happened to open that morning. Twenty years between Tesco, Dunnes Stores and Oracle taught me what happens when sensitive operational data leaks into someone else's training pipeline: it doesn't come back. So I built the thing I wished existed — an on-premise intelligence layer that lives inside the firm's own walls, learns the firm's own work, and answers to the firm's own people.
The problem I actually saw
By late 2023 it was obvious that staff in every Irish firm I knew were already using ChatGPT, Claude and Gemini for real work. Drafting client letters. Summarising contracts. Pulling numbers out of management accounts. Asking for second opinions on tax positions. The work was getting done faster, but the firm had no log of what was asked, no control over what was sent, and no way to keep the institutional learning from those interactions.
That's the bit that bothered me most as an engineer. A senior solicitor with thirty years of practice spends an afternoon iterating on a tricky probate question with a public LLM. The answer is fine. The reasoning is fine. But none of it stays in the firm. The next solicitor who asks a similar question starts from zero, with the same public model, and probably gets a slightly different answer. Multiply that by every regulated firm in the country and you've got a slow-motion knowledge leak that nobody is measuring.
The compliance side is the visible problem. Data Protection Commission, the Law Society, CCAB-I, the Central Bank — they all care, and rightly, about where client data ends up. But the deeper problem is that the firm's own intelligence is being externalised one prompt at a time, and the partners running these firms have no architectural answer to it.
Why I didn't just build a wrapper
The obvious thing to do — and what most Irish "AI consultants" are selling right now — is a thin wrapper around OpenAI's API with a custom system prompt and a Stripe checkout. I won't do that work. It solves nothing. The data still leaves the building. The firm still has no model of its own. The wrapper vendor goes out of business in eighteen months and the firm is back to square one with a pile of ChatGPT logs in someone's personal account.
What I wanted was an intelligence layer that satisfied four hard constraints at the same time:
- The model and the data both stay on hardware the firm controls. Either in their own server room or in a sovereign Irish data centre with a contract the partners can read in an afternoon.
- It learns from the firm's actual work — matters, files, correspondence, precedents, ledgers — not from a generic internet scrape.
- It produces an audit trail by default, so when the regulator asks "what did this system tell your staff and on what basis", there's an answer that takes minutes to retrieve, not weeks.
- It survives me. If I get hit by a bus, the firm still owns the model, the weights, the indexes and the runbooks. No vendor lock-in dressed up as a subscription.
That set of constraints rules out almost every "AI for business" product currently being pitched in Ireland. Which is why I built my own.
The architecture, briefly
The Intelligence Brain is three layers stacked on commodity hardware. The bottom layer is an open-weights base model — typically a Llama or Mistral derivative — chosen for the firm's hardware budget and language requirements. I don't train these from scratch; that would be wasteful and dishonest. I fine-tune and adapt them.
The middle layer is retrieval. The firm's documents — practice management exports, document management system contents, accounting backups, email archives where appropriate — get ingested, chunked, embedded and indexed into a vector store that lives on the same box as the model. Every answer the system produces cites the source documents it drew from. No citation, no answer. That rule alone kills most of the hallucination problem that public chatbots suffer from in professional contexts.
The top layer is the policy and audit harness. Every query and every response is logged with the user, the timestamp, the documents retrieved, the model version, and the policy decisions made along the way (was this query allowed for this user, were any documents redacted, was the answer flagged for review). That log is what a partner shows the Law Society or the DPC when they ask what the system has been doing.
The whole thing runs on a single well-specified server for most firms I've costed. It is not magic. It is engineering done carefully.
Why on-premise, specifically, in Ireland
I get asked weekly why I don't just put this on AWS Dublin or Azure Ireland and call it sovereign. The honest answer is that "Irish region" on a US hyperscaler is a contractual fiction in 2024. The CLOUD Act exists. The data residency promise is real until the day a US court order says it isn't. For a Tipperary solicitor handling a contentious probate, that's a risk worth avoiding entirely if the alternative is a server in their own building that costs less than a junior salary over five years.
The other reason is latency and offline tolerance. A rural Irish firm with a flaky broadband connection cannot have its primary work tool fall over because of a fibre cut between Cork and Frankfurt. On-premise removes that whole class of failure. The model answers at the speed of local RAM, not the speed of someone else's API gateway.
And finally — and this is the founder bit, not the engineer bit — Irish professional firms are built on relationships and trust that took generations to accumulate. Asking a sixty-year-old managing partner to trust a Californian company with thirty years of client correspondence is a conversation that goes nowhere. Asking him to trust a server in his own server room, configured by someone he can drive to in an hour, is a conversation that actually happens. I live in Clonmel. I can be in most Irish cities before lunch. That matters more than any technical argument.
What it actually does for the firm
The day-to-day value shows up in four places. First, drafting — letters, memos, file notes, first-pass contract reviews — done against the firm's own precedent bank rather than a generic model's idea of what an Irish letter of engagement looks like. Second, retrieval — staff asking questions across the entire matter history of the firm and getting cited answers in seconds rather than billing two hours to "review the file". Third, onboarding — new staff have a knowledgeable colleague available at three in the morning that has read everything the firm has ever produced. Fourth, partner-level oversight — the audit log gives partners a view of what their team is actually working on and asking about, which is something most professional firms have never had.
I deliberately don't pitch productivity percentages. I've seen too many AI vendors quote made-up numbers and I won't add to that pile. What I will say is that every firm I've installed for has, within the first month, found uses I didn't anticipate when I scoped the project. That tells me the underlying capability is real. The specific ROI is for each firm to measure on its own terms.
If you want the longer-form pitch with the vertical-specific detail, the Intelligence Brain overview page lays out how the same architecture adapts across legal, accounting, property, medical, education and public-sector deployments.
Why I'm doing this from Clonmel
People ask, gently, whether an Irish AI founder ought to be in Dublin or San Francisco. I understand the question. The answer is that the firms I'm building for are not in San Francisco. They're in Clonmel, Kilkenny, Waterford, Cork, Limerick, Galway, Sligo. The work is in the field, in their server rooms, sitting at their conference tables explaining what a vector index is to a managing partner who qualified in 1987. You cannot do that work from a co-working space in SoMa, and frankly, the engineering doesn't get any better for being done with a view of the Bay Bridge.
Tipperary is also, practically, a good place to run a software business in 2024. The fibre is fine. The talent that wants to live here is excellent and loyal. The cost base is sane. And the hardware sits in a building I can walk to in fifteen minutes, which matters when a customer calls at half nine on a Tuesday because the indexer has fallen over.
Where to start this week
If you run a regulated Irish firm and you've read this far, the most useful thing you can do this week is a five-minute audit: ask your staff, honestly and without blame, which AI tools they're already using for client work, and what data has gone into them. You will be surprised. That conversation is the start of every Intelligence Brain engagement I do, and it costs nothing to have. If the answers worry you, get in touch and we'll have a sensible conversation about what an on-premise alternative would look like for your firm specifically — no slides, no jargon, just the engineering.