We work on data nobody has validated a model on: a repertoire outside the training distribution, clinical outcomes from small cohorts, decades of one organisation's records. We build the evaluation set with your experts, rank the labels to check first, and hold part of the set back from everyone, us included. Every output records its evidence, and we test what a decision depends on by changing the input. The system runs on your infrastructure. The model is open-weight, fine-tuned on your material, and handed over with its evaluation sets. We work within the EU AI Act, ISO/IEC 42001, GDPR, the Swiss FADP and KVKK.
A Turkish music catalogue: microtonal modes (makam), asymmetric rhythmic cycles (usül, aksak). Audio stays on controlled infrastructure.
We built the analysis pipeline, tagging to similarity, with per-feature provenance, evaluated against an expert-annotated makam and usül set. In use; the attribution layer above it is current research.
Gait and IMU sensor data with open clinical corpora; patient data off third-party clouds. We built and demonstrated recovery-trajectory prediction with an industry partner. Concluded.
A tagging model looked accurate and read the accompaniment where the annotation named the singer. Attribution located the confound, removing that source from the input confirmed it, and the fix went into the training data. It also ranked the annotations to re-check and found one recording that was not the one its record named. The same audit applies to any classifier over audio, sensor streams or records; it is offered below.
A Horizon Europe research proposal on provenance and attribution for cultural content in the era of generative AI, with a European consortium; we lead the on-premise analysis and attribution work package. Under evaluation.
We also take advisory retainers and work-package roles in national and European research programmes. Write to contact@gurma.ai with a few lines about the data you hold.