AGENTIC AI
Anyone can demo an agent.
I run them in production.
Forty vendors will show you an agent booking a meeting. The question that matters is what it does at 3am when the API times out, the data is malformed and nobody is watching. That is the part I build: AI agents for businesses in Greece and across the EU that keep working in production.
WHAT AGENTIC MEANS HERE
Decides, not executes.
A script
Does exactly what it was told, in order, or dies. When the input shifts, it ships wrong answers quietly.
An agent
Looks at what came back, decides the next step, retries with a different approach, escalates to a human when it should.
The difference
State, tools, evals and cost ceilings: engineering, not prompts. That is what separates a system from a demo.
THE STACK
BUILT FROM SCRATCH
I built my own agent harness. Multiple times.
Once for cloud constraints, once for a local machine, because the two have nothing in common. Queues, retries, token budgets and observability in the cloud; latency, privacy and offline behaviour on the desk. Most people consume a framework. Building the harness twice is how you learn what the frameworks are hiding.
WHAT I BUILD FOR BUSINESSES
Agents with a job.
Ops agents that clear queues. Research and ingestion that reads what your team can't get to. Decisioning with a human in the loop where it belongs. Internal automation tied to a number someone actually reports on. Not capabilities, outcomes.
TRAINED, NOT PROMPTED
I did not call an API.
I trained the model.
Most AI work is a prompt in front of somebody else's model. BSLM is the other kind. Every weight starts from random: its own tokenizer, its own corpus, its own training loop. A 5M router reads the request, a 72M decoder decides what to do about it, in English and Greek, shipped as one 78 MB file that runs on a plain CPU or a phone. No API key, no per token bill, nothing leaving the building.
How it was made, in order: fix the ceiling at 80 MB so it fits a phone, generate the corpus, train an own byte level tokenizer, pretrain the decoder from random init on rented GPU time, then teach it to use tools from real trajectories rather than scripts, so it searches, reads, retries and judges its own steps. Thirteen measured rounds, and the ones that made it worse are published with the rest.
91.2% on held out agent tasks, 94.8% recovery after a failed first attempt, 2.8% wrong answer delivered without a warning. The one family under its bar, open fact lookup at 68%, is on the scoreboard rather than hidden.
See the code Download the model MIT licence, 78 MB GGUF, runs on a CPU
WHEN IT GOES WRONG
Anyone can build an agent.
Few can keep one alive.
The architect section, because this is where the demos die:
Retries & fallbacks
Every tool call fails eventually. The system's answer to that is designed, not discovered.
Evals
A model update can silently change behaviour. Evals catch it before your customers do.
Cost ceilings
An agent in a loop is a credit card in a loop. Budgets are enforced in code, not in hope.
Human in the loop
Some decisions should never be autonomous. Knowing which ones is the job.
I ALSO TEACH IT
learnery.io
AI training your team will actually finish: a five-step, hands-on seminar teaching your people to use AI tools confidently, tailored to your industry and their exact roles. Teaching a subject is the strongest proof of expertise in it, and some teams need the capability in-house, not delivered.
See learnery.ioALREADY SHIPPED
AI in production before 2026.
seethemenu.online has shipped an AI assistant to real businesses since before the pivot year. learnery.io is an AI product live right now. The claim on this page has receipts that predate the hype: URLs, not adjectives.
DATA & GDPR
Where your data goes,
answered before you ask.
EU posture by default: data minimisation, EU-resident processing where the stack allows it, retention that is decided rather than accidental, and a written answer to "what does the model see". I have built GDPR-compliant systems since 2018, when the regulation began to apply. The agent work inherits that, it does not escape it.
QUESTIONS BUYERS ASK
AI agents for business: what buyers ask first
Who builds AI agents for businesses in Greece?
I do, from Athens, for businesses in Greece and across the EU, meeting in person or working remotely. 12 systems I built already run, each with a dashboard you can open on this site, and customers use them. I have built for businesses in Greece for twelve years, and the first call is free. See them running
How much does an AI agent cost?
It depends on your volume and on what a mistake costs you. The written blueprint prices it before anything is built, counting the model calls, the retries and the review time a person still spends. Read more
How long until it is live?
A pilot on your own data comes first, sized to one narrow task, usually in weeks rather than months. What goes live afterwards is what the pilot proved, with the numbers written down.
Is my data safe?
The agent gets the narrowest access that does the job, personal data is minimised before it reaches a model, and providers are chosen for their data terms, under a processing agreement. Anything that cannot be undone waits for a person.
Chatbot or AI agent?
A chatbot answers. An agent acts: it reads what comes back from your systems, decides the next step, retries and asks a person when it is unsure. Many businesses need the first, some need the second, and sometimes the honest answer is a plain script. Read more
What does the EU AI Act require?
Customers must be told they are talking to an AI, and staff who use AI need literacy measures fitted to their role, both already in force. Agents that screen job candidates or score credit are high risk, with obligations from 2 December 2027.
BUILT, NOT PITCHED
12 systems already built,
in use by customers.
Tax reconciliation, agent governance, EU AI Act evidence, agentic checkout, nonprofit marketing, retail stock and cost, customer support, HR screening, sales and logistics. Complete and running.
See the 12Have a process that should run itself?
Bring the process and the number it burns. I'll tell you on one call whether an agent is the right answer, including when it isn't.
Start a project