Meet Ecobot: Eco Vent’s New AI Ventilation Assistant — and an Honest Look at the AI Behind It

You may have already noticed it. In the bottom-right corner of ecovent.ie, there is now a small chat bubble. That’s Ecobot — our new AI assistant, and the first thing I want to tell you about in this post.

Ecobot is your first line of knowledge about mechanical ventilation and about Eco Vent itself. Ask it a question — about MVHR, MEV, DCV, MEV, about grants, about how a ventilation system is operating, or simply about what we do — and it will read through our website and give you an answer in a few seconds. No waiting, no phone tag, no office hours.

But because I built this little assistant myself, I think you deserve more than a marketing line. I want to tell you what’s actually under the bonnet, what AI can and can’t do, and how a one-person ventilation business in Co. Laois ended up running on it. So this is part announcement, part honest field report.

Before any of this, I had to learn how to be found at all

Long before ecovent.ie became one of the best-ranking ventilation websites in Ireland, there was a much more basic problem: nobody could find it.

To fix that, I had to teach myself digital marketing more or less from scratch — SEO, power words, internal linking, mobile-first indexing, and the way Google rewards pages that load fast on a phone (these days that’s measured as Core Web Vitals). Altogether it took me around three months of research and work, and hundreds of hours getting to grips with the rules that actually govern how websites rank.

Honestly, without AI — and especially Perplexity’s deep-research mode — I think I’d have spent years going round in circles, if I hadn’t shut the business first. Marketing is, by many accounts, one of the fields AI is now reshaping fastest, particularly the junior roles. That cuts both ways: the same tools that put pressure on that profession are the ones that let a sole trader like me do the work himself.

And that matters, because here’s the offer I got early on:

“When my site was still barely visible online, a marketing agency offered to promote my business. They asked how many visits I was getting a month. Ninety, I said. For €2,300 a month, they told me, I’d get forty more. I thought it had to be a joke.”

“After three or four months of hard work on my own, the result blew past my wildest expectations: over 800 visits a month.“

The skills I picked up give me something more valuable than traffic, though — they give me the freedom to steer my own business. My marketing won’t be sales gibberish. It’ll be an answer to the real needs of the people I meet on the job.

Here’s one of those needs. Looking back over the systems I’ve worked on, proper documentation left for the homeowner is a rarity — a record of how to use the system, that it has to be serviced, and least of all a commissioning report. I could count on one hand the installations I’ve come across that had documentation like a commissioning report left behind.

“In situations like that, it’s always the client who loses out in the end.”

Measuring and balancing airflows is a service I charge for — fair enough. But it was also the installer’s duty: to actually balance the system properly in the first place, and to leave behind a record proving it was done. The Android field-service (FSM) app I’m building with Claude is aimed squarely at this — making that documentation and handover a standard part of every job, for my sake and the client’s.

First, what is an LLM — in plain English

Ecobot is built on what’s called a Large Language Model (LLM). The name is worth unpacking.

• Large — these models are trained on an enormous amount of text and have billions of internal “settings” (parameters).
• Language Model — at its core, the model does one deceptively simple thing: it learns to predict the next word in a piece of text. Do that across a vast amount of writing and, surprisingly, the model starts to “understand” grammar, facts, reasoning patterns and tone.

That’s it. The name comes from modelling language — not from speaking many languages. The popular AI tools you may have heard of — ChatGPT, Claude, Gemini — are all LLMs in exactly this sense.

So did these models really teach themselves other languages on their own? Sort of — and this is the genuinely interesting part. The training data (huge crawls of the internet) did contain text in many languages, but English dominated massively; Polish, Irish, Dutch and the rest were a small fraction. What’s remarkable is how much capability the models squeeze out of that small fraction, and how knowledge learned in English “carries over” into other languages — researchers call this cross-lingual transfer. A model can translate, or answer a Polish question, even though no one explicitly sat down and taught it to do exactly that.

The multilingual ability is a genuine, emergent surprise — it just isn’t where the letters “LLM” come from.

And yes, English really is over-represented in the training data. That’s not a rumour; it’s a fact about how these systems were built. Which leads me neatly to my own story.

What AI changed for a one-man business

I’ll be straight with you: English is not my first language. Writing it cleanly, professionally, without obvious mistakes — from where I started, that would simply not have been possible without AI helping me with the text. I built the Eco Vent website myself, with AI as my writing partner.

Here’s the funny part. For a long time I wrote my prompts in English, even though Polish is my native language — because I often didn’t know the Polish word for a piece of ventilation kit, but I knew the English one. And when I was building the first Eco Vent pages, I genuinely had the impression that a prompt written in my broken English produced a better result than the same prompt in polished Polish.

“I’d type the prompt in clumsy English and still get a cleaner answer back than if I’d written it perfectly in Polish.”

That impression was well-founded — it was the English-in-the-training-data effect I mentioned above. The interesting thing is that the gap has now largely closed. Three or four years ago it was very noticeable. Today, the difference between languages is hard to spot.

Is the website just “made up by AI”? Absolutely not.

This matters to me, so let me be clear. The content on ecovent.ie is not an AI invention dressed up as expertise. The AI is a language model — it’s brilliant at putting things into words — but the knowledge, the standards, the regulations and the judgement are mine.

And here’s where AI showed its limits. I repeatedly ran into trouble with technical terms that look similar but mean different things — for example, the difference between validation, a validation certificate, and a commissioning certificate/report. The model would happily confuse them. The more technical and specialised the language got, the more the model tended to “fantasise” — to fill gaps with plausible-sounding but wrong statements. The last few technical articles for the site were genuinely hard work for exactly this reason.

What actually produced good results

A few hard-won lessons:

• A very precise prompt. Spell out what’s allowed, what’s not allowed, and where the knowledge should come from (e.g. the relevant Building Regulations PDF, S.R. 54:2014, manufacturer data). The single biggest quality jump comes from being specific about what you want at the end.
• Mind the context window. Every model has a limit on how much it can hold in its “short-term memory” (input + output tokens). Push past it and the quality drops, mistakes creep in, or the model stalls completely. The fix is discipline: break the work into shorter, focused sessions instead of one giant marathon.

Research superpower — with a catch called hallucination

AI is an outstanding research tool. But it can also confidently invent things — a “hallucination.” Two practical defences I rely on:

• Source-linked answers. Moving a lot of my research from a plain chatbot to Perplexity, which returns links to its sources, was a turning point. If I can click through and check, hallucinations have nowhere to hide.
• RAG models — answers grounded in your data. RAG (Retrieval-Augmented Generation) means the AI answers using a specific set of documents you give it, rather than its general memory. The hallucination risk drops sharply. Google’s NotebookLM is a great example — point it at your own files and it largely sticks to them. Ecobot itself works on this principle: it answers from our website, not from the open internet.

A few tips, especially if you’re just starting with AI

• Write a very precise prompt. Put special emphasis on the end result you want — the format, the tone, the length, what to include and what to leave out.
• Different tools are good at different jobs. Don’t expect one model to be best at everything. Here’s my own, personal take after a lot of daily use:
  • ChatGPT — strong on creativity and reasoning, and excellent at voice and live conversation. (I had the impression it was among the first to actually modulate its voice — change tone and warmth — mid-conversation.) The downside I’ve felt lately: it’s the most likely to ignore my instructions and do its own thing.
  • Perplexity — my go-to for research on the web. It finds pages you simply won’t surface through a normal Google search, and it lets you switch between several underlying models. Weaker on deep reasoning and text analysis, but for fast, sourced research it’s invaluable.
  • Switching between tools depending on the task is the real skill.

What I actually pay for

Right now I run two paid AI subscriptions:

• Perplexity — indispensable when I need to find the manual for a specific ventilation unit fast, without losing half an hour to it.
• Claude — currently my number one. I’m even looking at moving up a subscription tier, because I’m slowly bumping into my usage limits. It’s superb at analysing text, genuinely excellent at writing code, has a great desktop interface where I can watch how it’s “thinking,” and it’s very disciplined about following instructions. On top of that there’s a whole range of connectors — apps that exchange data, automations, and an “agent” mode whose potential I’m only beginning to tap.

“I already know this much: this company won’t need a secretary.”

Eco Vent is, right now, in the middle of automating every process involved in running a small business.

Back to Ecobot — how it was actually trained

Ecobot was trained on 48 articles — roughly the number of pages on ecovent.ie. And here’s the part I care about most:

Every single one of those articles was read by me, Lucas, to make sure the bot isn’t fantasising.

That’s the whole point. By grounding it in content I’ve personally checked, I keep it honest. I won’t pretend it’s perfect — some bugs can happen, and if you catch one I’d genuinely like to know. Under the hood, Ecobot is really two AI models that activate depending on context — one to understand and route your question, one to compose the grounded answer.

So when you ask it something in that little corner of the screen, you’re talking to a system built on knowledge I stand behind.

Try it yourself

Ecobot is live now, bottom-right corner of ecovent.ie. Ask it about your ventilation system, about grants, about booking a visit — or just test it with a hard technical question and see how it does.

And if you’d rather talk to a human, that human is still me. The AI just makes sure I have more time for the work that actually matters: getting your ventilation right.

— Lucas, Eco Vent