Tallinn, Estonia
Answers from your own records, without the records leaving.
We build retrieval systems that run inside your own infrastructure. Staff ask questions in ordinary language and get answers with the sources attached. Nothing is sent to an outside service.
The problem
The information is usually already in the building.
Most of the organisations we work with are not short of data. They are short of a way to reach it. The answer someone needs already exists, in a system three departments away, in a format nobody there can query. Generative AI is an obvious fit for that problem, and for organisations holding classified, personal, or commercially sensitive records it is usually also unavailable, because sending that material to an external interface is a disclosure whatever the supplier's terms say.
Government and defence
Case files, reporting, and operational records held across agencies that were never designed to be queried together, under classification rules that make an external service impossible.
Healthcare and public health
Clinical and administrative records under data protection obligations, where the person who needs the answer is a clinician and the delay is measured in patient outcomes.
Logistics and industry
Freight, port, and supply chain operations where the relevant history sits in a decades-old system that the people making today's decisions cannot search.
How it works
Three properties, each one a design decision we can show you.
Retrieval-augmented generation, deployed on infrastructure you control. The claims below are checkable, and we would rather demonstrate them than assert them.
The data stays where it is
Indexing and generation both run on hardware you control, and no part of the retrieval path makes an outbound connection. Search runs on ordinary processors, because most secure environments do not have graphics hardware and requiring it would put the system out of reach.
Access limits live in the database
Who may see which records is enforced by the database itself rather than by application code, so a change to the application cannot widen it. We verify this with an automated test and will show you the result rather than describe it.
Answers say where they came from
Every answer carries its sources. A query that finds nothing says so instead of estimating. We measure how often the system breaks its own rules and report that number rather than quietly correcting it.
Evidence
Where it runs today.
A reference deployment runs on PostgreSQL with the pgvector extension in Frankfurt, using hybrid retrieval that combines vector similarity with keyword search. A user holding the highest clearance level who belongs to a different client organisation retrieves nothing at all, and that is confirmed by a test rather than asserted on this page.
The same design is in pilot with a premier university athletics programme in the United States, covering medical, scouting, compliance, and market records. The client is not named for confidentiality reasons.
What we will show you
For anyone evaluating this seriously: the deployment itself, the access-control model, the isolation test output, and the written record of a retrieval fault we found during verification and the reason it was not visible from outside the system.
Working with us
An engineer comes with the software.
The model is rarely the difficult part. The difficult part is a registry that uses one set of regional codes and an employment dataset that uses another, or a records system nobody has queried directly in eleven years. That work does not happen over email.
So the core system is fixed and identical for every client, and one connector is written for yours. Over time we train engineers to that same standard, so delivery does not depend on any one person's calendar. That is our plan for growing, and it is deliberately a slow one.
Who you would be working with
Jonathan Trippett, MD, BCMAS. A physician and United States Army veteran who builds these systems personally, in code. Previous work covers field medical records that have to survive with no network at all, athletics intelligence, and public economic data.
The relevant experience is not artificial intelligence. It is having worked where mishandling sensitive information harms someone.
- Storage and search. PostgreSQL with the pgvector extension. Your organisation already knows how to back up and secure PostgreSQL, which means we are not introducing a second system for somebody to accredit. Records, search indexes, and access rules stay inside one boundary.
- Models. Open-weight models running on your hardware. Larger models for work that requires reasoning; small models are appropriate for narrow classification behind a deterministic check and not for decisions. Which model runs where is written down rather than left to configuration.
- Connectors. Written per client, reading only the systems named in the engagement. Every stored record traces back to the connector run that produced it, so the question of where an answer came from is answerable down to the source.
This page loads nothing from a third party. No external fonts, scripts, analytics, or tracking. It seemed the least we could do.
Contact
If you have a problem in this shape, we would like to hear about it.
We are taking on a small number of first engagements in Estonia at reduced cost, in exchange for a reference. The reason is straightforward: we are new here and would rather earn a local track record than claim one. A first project is one narrow problem, one data source, and a fixed end date.