No accuracy promises upfront
We do not quote an accuracy figure before seeing your data. Anyone who does is quoting someone else’s benchmark.
AI product studio · Software · Evaluation · Cyprus
Services
Four kinds of applied AI work and the four questions we ask before agreeing to any of them.
01 — What we build
Four kinds of AI work
All of it applied to a specific business process with a measurable before and after. None of it sold on the promise that the technology is impressive.
Document processing
Extracting structured data from invoices, contracts, specifications and the assorted PDFs a business drowns in. Measured against a labelled set from your own documents, not a public benchmark.
Classification & routing
Sorting incoming work — tickets, applications, messages — to the right queue. Usually the highest-value AI application in a company and the least glamorous.
Assistants with guardrails
Internal assistants that answer from your own material, cite where the answer came from, and say "I do not know" instead of inventing. That last property is engineered, not hoped for.
Evaluation harnesses
The part almost everyone skips: a repeatable test set that tells you whether a prompt or model change made things better or quietly worse.
05 — The boundary
What we do not do
Four limits, stated before you spend a call finding them.
No training on your data
Client data is not used to train shared models. The processing arrangement is written down before anything moves.
No legal advice on the AI Act
We flag where a use case looks higher-risk and recommend a lawyer. We are engineers.
No autonomy where errors reach people
Customer-facing decisions get a human in the loop. That is a design rule, not a phase we remove later.
02 — Approach
How an AI project runs
Evaluation first. It is the cheapest stage and the one that most often ends the project early and correctly.
Phase 01
Baseline
How is the task done today, how long does it take and how often is it wrong? Without this, no later claim of improvement means anything.
Phase 02
Eval set
A few hundred labelled examples from your real data, including the awkward ones. This is the honest part of the project.
Phase 03
Build
The smallest system that beats the baseline, with a human review step wherever an error would reach a person.
Phase 04
Watch
Monitoring in production, because model behaviour drifts and a system nobody watches degrades quietly.
In detail
Four questions we ask before building anything
If the answers are unsatisfying, the correct outcome is a short invoice and no project.
Question 01
What does wrong cost?
An error in an internal draft is cheap. An error in a customer-facing decision is not. The answer determines how much human review the design needs.
Blast radiusReview depthReversibility
Question 02
What is the baseline?
Time per case and current error rate. Nobody measures this before an AI project, which is exactly why so many are declared successful without evidence.
TimeError rateVolume
Question 03
Can we label the data?
A few hundred examples with correct answers. If nobody in the business can produce them, nobody in the business can tell whether the system works.
LabelsEdge casesOwnership
Question 04
What happens when it is down?
A model endpoint will be unavailable at some point. If there is no fallback, you have introduced a dependency rather than a capability.
FallbackDegradationAlerting
04 — FAQ
Clearly answered
Will you tell us AI is the wrong tool?
Regularly. A rule engine or a better form is often cheaper, more predictable and easier to defend. That answer costs us a project and saves you one.
Whose data trains what?
Yours stays yours. We do not use client data to train shared models, and any processing arrangement is written down before data moves.
What about the EU AI Act?
Most internal document and classification work is low-risk, but the obligations depend on the use case. We flag where a use case looks like it could fall into a higher-risk category and recommend legal review. We are engineers, not lawyers.
Can you improve an AI feature we already have?
Yes, and the first step is building the evaluation set that should have existed. Without it, nobody can tell whether a change helped.
Do you host models?
We integrate whatever fits — hosted APIs or self-hosted models. The choice depends on the data sensitivity and the cost profile, not on a preference.
Which markets do you serve?
Registered in Cyprus, delivering across the EU, in English and German.
Let’s talk
An AI project that needs an honest second opinion?
Tell us the task and how it is done today. If the answer is that you do not need a model for it, that is what you will hear.