I started Leemer Chat in February 2023 because GPT-4 went down and I needed somewhere to keep working. That was the whole reason. It became the group's first product, and it is still the one that taught me the most about what the rest of them would need.
We are five products now: Leemer Chat, Leemer Labs, Critique.sh, DáilDex, Founder2, alongside Square Pigeons, the studio that builds software for other companies. Between them they have served thousands of people, passed more than a hundred billion tokens through one of them, and now carry work from people who write code for a living, students in Ireland, and members of the Irish parliament's own offices.
People ask how a small group in Waterford runs that. It is not because we are unusually clever and it is not because we have secret infrastructure. It is because we standardised on a handful of habits, and then refused to let a product grow past the habits that made it worth using.
A group gives separate products three things: shared research, shared engineering and shared credibility. It also gives them three risks: shared failure, shared distraction, and the temptation to make five products look like one platform.
We took the first two and refused the third. Each product keeps its own name, its own interface and its own audience. A developer reviewing a change should not have to learn a civic data product's interface, and a person following what their TD did this week should not have to care about pull requests. Making one product's interface the template for the others would have made both worse.
So the thing we standardised is not the interface. It is the way we build.
Leemer Chat is the multi-model workspace the group started with, and the place where a new frontier model usually lands first, before it becomes a product anywhere else.
Leemer Labs is the research and infrastructure arm: model training, fine-tuning, evaluation and inference, hosted in Ireland. It trained Born2, a nine-billion-parameter code reviewer on a fifty-billion-token dataset of our own, and it was the first organisation in Ireland to work with Thinking Machines' Tinker in closed beta.
Critique.sh is the one that came out of a specific frustration: an agent writes a change quickly, and then somebody still has to decide whether the change is sound. Producing a change and establishing that it is sound are different jobs. So Critique reproduces the change in a clean workspace, tries to break it with a different model family, and hands back a verdict with the evidence attached.
DáilDex makes Irish parliamentary information followable: follow a TD, get plain-English updates with the source linked, ask a question and get the answer traced back to the record. It is our fastest-growing product, it is free, and it is used by members of the Oireachtas, by people working in the offices of the Minister for Health and the Taoiseach, and by anyone who has ever wondered what their representative actually did.
Founder2 is the newest: a workspace for a solo founder running a SaaS company, with a living plan, a task board, connected sources — Stripe, GitHub, Sentry, PostHog, Slack — and a co-founder that works against that state. It is the product this journal belongs to, and it is built on the same habits as the other four.
100bn+
Tokens through Leemer Chat
Around 80% of them in 2026 alone.
234
Representatives in DáilDex
Across 43 constituencies.
Hundreds of millions
Lines of code reviewed by Critique.sh
Billions of tokens since launch.
5
Products from one group
Waterford, Ireland.
Leemer Labs exists because we wanted to know why one model behaved differently from another on the same task. That question turned into an evaluation lab, and the evaluation lab turned into the review engine behind Critique.sh, and the review engine turned into the way we talk about our own products.
The rule is simple and slightly awkward to live with: a confident summary is not evidence. When Critique says a change is ready, it hands back a reproduction where it can, the checks it ran, and an explicit account of what remains uncertain. When we publish a number, we can say where it came from. When we publish a capability, we can point at the code.
For a product serving hundreds of users, this is not a stylistic preference. It is the difference between a user who trusts the tool on the one morning it is wrong, and a user who never trusts it again.
Every agent we ship runs somewhere it cannot see everything. In Founder2, the co-founder works in a sandbox with its own filesystem and its own credentials, and it reaches the workspace through an authenticated MCP channel. It cannot read the database directly. It cannot read another workspace. Every write to an external system — creating a GitHub issue, commenting on one — is held until a person approves it, with the destination and the proposed wording visible first.
This came directly out of Critique.sh, where the whole product is the idea that an agent should not get to declare its own work correct. If we are going to insist on that standard for a pull request, it would be strange to be lax about it inside our own workspace.
It also does something for cost and safety that is easy to miss: an agent that can only act through a narrow channel is an agent we can reason about. We can answer "what could it have touched?" without a spreadsheet.
We are an Irish company, and the group runs its own inference in Ireland. That is partly practical — latency and cost — and partly a decision we made early and have not had to revisit.
In practice it means several things. The application and its database run on European infrastructure. Model traffic goes through OpenRouter, which supports EU in-region routing on its Business and Enterprise plans, and can be pinned to European endpoints so that a request either stays in Europe or fails rather than quietly leaving. Connected sources can be pointed at their EU regions. Our own models — Born2 and its successors — are trained and served here.
None of this makes us compliant by default, and we do not say that it does. It makes compliance a property we can demonstrate rather than a hope. A customer in Dublin or Eindhoven or Munich can ask us where their data goes, and the answer is specific instead of reassuring.
The hardest habit in this list is also the one that has the most direct effect on whether people keep using what we build. We name what does not exist.
Not every integration is implemented. Not every connected source can be written to. Some actions need approval and some do not. Some models are good at a task and others are not. A marketing page that implies a capability exists is a small lie that compounds: the person who discovers it tells everyone they know, and the next thing you say is discounted.
So the journal, the product pages and the sales conversations all carry the same sentence: this is what is built today, and here is the boundary. It costs us conversions we would rather have. It is also the reason people come back a second month.
Serving hundreds of users does three things to a product that serving one does not.
It removes the personal correction loop. When one person uses a tool, you learn what they wanted from what they did next. At hundreds, you need the tool to teach you, which means the decisions have to be legible in the interface rather than in a support conversation.
It makes every sharp edge systemic. A bug that one user works around is a bug for one user. A bug that a hundred users work around differently is a design decision you have not made yet. Our own review tooling caught this early: the same mistake kept arriving in different codebases, and the only fix that worked was removing the opportunity, not explaining the rule.
And it makes trust a performance property. Leemer Chat has passed more than a hundred billion tokens. Critique.sh has reviewed hundreds of millions of lines of code. DáilDex is used by people whose professional reputation depends on the record being reported accurately. Each of those numbers is a promise about the next request, and the next request is the only one that matters.
Write the boundary down before you write the feature. Decide what the product will not do, publish it, and treat it as a contract. The products that survive contact with hundreds of users are not the ones with the most features; they are the ones whose claims stayed true.
Then automate the part you can check and keep the part you cannot. In our group that split became Critique.sh, and it is the reason we can let an agent loose on a codebase without lying to anyone afterwards. The same split is inside Founder2: the co-founder does the drafting, the reading and the arithmetic, and the decisions that change something outside the workspace still belong to a person.
Choose where you are hosted before you choose where your model is. It is much harder to move a data boundary later than to set one at the start.
We have been wrong about the shape of this more than once. Leemer Chat was meant to be a backup and became the group's flagship. Critique.sh was meant to be a side project and is now the product with the hardest credibility bar. We were early to believe that agents would sit inside workflows rather than beside them, and early to believe that open-weight models would close the gap for a specific task as quickly as they did.
The pattern in those errors is the same: we underestimated how much the interface, not the model, decides whether a tool gets used.
We are still wrong about some of it. Founder2 is early, the list of connected sources is short, and the honest answer to "how much of this is built" is: a useful amount, with edges that we describe rather than hide.