93%
It finds the right law.
Before answering, Merkys looks up the relevant legal text. In 93% of cases it finds every passage the answer needs.
How we work
How we work.
How the products are deployed, how we measure them, and the research they come from.
Deployment
Why on-premise matters.
All three products run on the institution's own servers. No patient data leaves the building. No third-party cloud.
This isn't a feature — it's the foundation. Lithuanian healthcare data stays in Lithuania.
- Local-only mode
- Tamper-evident audit log
- Access control per document
- Automatic data encryption
Numbers
Numbers, not promises.
Few AI vendors publish how well their system actually works. We measure ours and publish the results for the current version.
Current version of the Merkys legal assistant: 116 legal questions, answers checked by people and by a separate AI model outside our systems.
41/50
41 of 50 legal questions answered correctly.
In the other cases it mostly gave a partial answer or said the information was missing, rather than making something up.
150/150
Every claim has a source.
In all 150 answers, every number and every reference matches real legal text. Checking by hand, we found one case where the model relied on its own knowledge instead of the text.
21/25
21 of 25 contract clauses assessed correctly.
In the other 4 the model didn't approve the clause because it wasn't sure. That's the safe kind of mistake: a person takes a second look.
25/25
When there's no answer, it says so.
We asked 25 questions with no answer in the documents. Every time, Merkys said it didn't have the information instead of inventing one.
What we're improving next.
Finding more
Adding more of the laws that healthcare acts refer to, so the right passage is found more often.
Checking harder
A second check that legal terms aren't confused, run on a different AI model for independence.
Measuring better
A second independent reviewer, and tests with deliberately planted mistakes to see how many it catches.
Research
Grounded in research.
Our products come out of AAI Labs' own research and are checked with internal benchmarks.
All our research →- 01
Doctor Assistant
is part of a larger European research project (InoMedicine) on Lithuanian medical speech recognition and record summarisation.
- 02
Merkys
builds on our research into language models that answer accurately and don't make things up, measured with the benchmarks above.
Talk to us.
Tell us which product interests you, and we'll get back to you.