We work where generic AI fails: on data that is proprietary, sensitive, or culturally particular enough that off-the-shelf models get it wrong, and confidential enough that sending it to a third-party cloud is not an option.
Our systems are built to run on the owner's own infrastructure, trained only on material the owner has authorised, and designed from the outset so their decisions can be inspected rather than trusted on faith.
Systems run on the client's own infrastructure. Recordings, patient data and proprietary records do not go to third-party AI clouds. An architectural commitment made before a project starts, not a configuration option added at the end.
Mainstream models are trained on the centre of the distribution. Turkish makam repertoire, clinical outcome trajectories and decades of one organisation's operational records all sit outside it. We build for that gap — and measure whether we closed it.
Every system is designed so its decisions can be inspected: what evidence produced this output, how confident is it, and where does it fail. Interpretability is the deliverable, not a feature added for a compliance review.
We work across the EU AI Act, GDPR and KVKK, medical device regulation, and the European and Turkish research funding frameworks. Compliance shapes the architecture from day one rather than arriving as a blocker.
Gurma AI AG is founder-led and deliberately small. The team combines more than twenty years of software and machine-learning engineering with active research in mechanistic interpretability — and both founders are practising musicians with working familiarity with Turkish makam traditions, which is why the cultural layer of this work is accurate as well as the technical one. Engagements are taken one at a time; founder profiles are available on request.