What the Intelligence Brain actually is
The Michael English Intelligence Brain is an organisational intelligence layer that runs inside your network. It is not a chatbot. It is not a wrapper around a public model. It is a coordinated set of language models — a swarm — that read your documents, your policies, your case files, your matter notes, and your internal procedures, and answer questions against that material with citations.
The technical shape of it: local model weights, a private vector store, a retrieval pipeline against your own documents, an orchestration layer that routes a query through several specialist models, and an auditor model that checks the output before it reaches the user. Everything stays on hardware you control. Nothing leaves the building unless you decide it does.
That is the precise definition. It is plumbing. It is the thing that lets a fee-earner ask "what did we tell the client in March about the indemnity clause" and get a real answer with the source paragraph attached, without that question travelling across the public internet.
The two markets it's for
I built this for two specific buyers. I am not chasing everyone.
Regulated professional services
Solicitors, accountants, insurance brokers, financial advisers, medical practices, and the consultancies that work alongside them. These firms have three things in common: a regulator who cares where client data lives, a professional indemnity policy that asks awkward questions about AI use, and a partner group that has been told by someone they trust never to paste client matter into a public chatbot.
For these firms the public-cloud AI products are a non-starter on day one. The conversation never gets past the data-protection officer. An on-premise layer changes that conversation.
Irish SMEs of 10 to 200 staff
The second market is the Irish mid-sized firm. Manufacturing, distribution, professional services, family businesses with two or three generations of process locked inside the heads of long-serving staff. These firms have the same productivity problem as everyone else, but they have a sharper version of the institutional knowledge problem — when the operations manager retires, twenty-five years of how-we-do-it walks out of the building.
I picked these two markets because they are where I have actually worked, and because they are underserved. The hyperscalers chase enterprise. The SaaS crowd chases startups. The middle ground — a sixty-person solicitor's office in Cork, a hundred-person manufacturer in Limerick — gets a generic product and a forum.
What "on-premise" really means in 2026
On-premise has become a marketing word. I want to be precise about what it means here.
The Intelligence Brain runs on hardware that sits in your office, your server room, or a data centre rack you lease and control. Model weights are local. The vector index is local. The retrieval pipeline is local. The orchestration layer is local. When a user asks a question, the question, the retrieved context, and the answer all stay inside your network boundary.
The network architecture is straightforward. The Brain sits behind your firewall on a segregated subnet. Users reach it over your internal network or via your existing VPN. There is no outbound dependency on a third-party API for inference. There is an outbound channel for software updates, which you control and can air-gap if your posture demands it.
The trade-offs are real and I will name them. You buy or lease GPU hardware. You accept that the model you run today is not the absolute frontier model that the hyperscalers ship next quarter — though for document retrieval and structured reasoning over your own corpus, the gap matters far less than the marketing suggests. You take on a maintenance relationship, either with my team or with an internal IT function I train.
What you get in exchange is a posture you can defend to a regulator, an auditor, and an insurer in plain English. Your data does not leave the building. That sentence either matters to you or it does not.
The audit-trail-first design
Most AI deployments treat the audit trail as a feature you bolt on later. I treat it as the foundation.
Every query into the Brain produces a record: who asked, what they asked, which documents were retrieved, which models were involved, what each model contributed, what the auditor model said about the draft answer, and what the final output was. That record is immutable, timestamped, and queryable.
The auditor is the most important model in the swarm. Its job is not to draft, summarise, or extract. Its job is to read the proposed answer and the source material and decide whether the answer is supported by the source. If it isn't, the auditor flags it, and the user sees the flag. The auditor runs against a tighter rubric than any of the drafting models. It is the smartest one in the room because being the smartest one in the room is the job.
This matters for a regulated firm because when the question comes — from a client, from a regulator, from a court — about how a particular answer was produced, you have the record. You can show the source documents, the chain of reasoning, the auditor's check, and the final output. That is the difference between a defensible AI deployment and a liability.
The shadow-mode rollout
No swarm of mine is allowed near a live tool until day 90.
For the first ninety days the Brain runs in shadow mode. It sees the queries. It produces the answers. It logs everything. But its outputs are not shown to end users in the workflow. They are reviewed — by me, by your nominated subject-matter experts, by the people in your firm who actually know what the right answer looks like.
The reason is simple. Until I have watched the Brain answer your questions against your documents for ninety days, I do not know where it is wrong. Neither do you. Shadow mode finds the gaps in your document corpus, the ambiguities in your internal policies, the queries the swarm handles well, and the queries it does not. We tune, we add sources, we adjust the auditor's rubric.
On day 90 you decide whether to go live, stay in shadow longer, or stop. I would rather lose a deployment at day 90 than ship a tool that gives a fee-earner a confident wrong answer in week one.
What the free assessment actually involves
It is a thirty-minute call. There is nothing to install, no data to share, and no follow-up sales sequence.
What I cover on the call:
- The shape of your firm — headcount, regulator, where your documents live, what systems your people use day to day.
- The specific problem you are trying to solve. Knowledge retrieval. Policy lookup. Drafting assistance. Onboarding. Whatever it is.
- Whether on-premise is actually the right answer for you. Sometimes it isn't, and I will tell you that.
- A rough sense of the hardware footprint and the timeline a deployment would involve.
- The honest constraints — what the Brain does well, what it does not do, and what is still hard.
What you walk away with: a written one-page summary of what I heard, what I would suggest, and whether I think you should do this at all. If the answer is not yet, or not with me, I will say so. The assessment is free because the work of figuring out whether a deployment makes sense is worth doing properly, and putting it behind a paywall would select for the wrong customers.
Frequently asked questions
Is the Intelligence Brain on-premise or cloud?
Default is on-premise — the firm's own server, the firm's own data, the firm's own model weights. We support private-cloud (your AWS, your GCP, your Azure tenant) when on-prem hardware isn't a fit. We do not run a multi-tenant SaaS.
What models does the brain run?
It depends on the vertical and the data-residency posture. We run frontier models (Claude, GPT) for high-stakes reasoning when the regulatory regime allows, and fine-tuned open-source models (the Qwen, GLM, and DeepSeek families) for everything else. The auditor agent is always the highest-capability model in the stack.
How long is the rollout?
About six months from kick-off to live use. Four eight-week stages — ingest, structure, swarm, audit. The swarm runs in shadow mode for the first ninety days alongside your team; only at day ninety, with the audit logs to back it up, does the swarm earn the right to run a tool live.
What does it cost?
Per-firm engagement, scoped from a free thirty-minute assessment. Firms vary too widely for a public list price — a five-partner law firm and a forty-person SME need different scoping. Book a slot via Calendly and we will scope it together.
Who am I talking to?
Michael English. I am the founder, the architect, and the engineer on every engagement. I do not run a sales team. The 30-minute assessment is with me directly.
Do you work outside Ireland?
Yes — Ireland and the UK are the primary markets. We have done methodology work for EU firms via Anglophone partners. The on-premise architecture means jurisdiction is decided by where the server lives, not where I do.
Can it write contracts / draft accounts / produce clinical letters automatically?
It can produce a first pass that a qualified human reviews before anything is signed, filed, or sent. Tool-layer authorisation is a hard architectural boundary in the brain — the swarm reads everything and signs nothing.
What about hallucination?
The auditor agent's job is to catch hallucination before output reaches a human. Every claim in every output is required to be cited; every cite has to be reachable; every cite has to load. If the auditor cannot verify, the output is rejected as a build-failure signal — not corrected.