Hard cases

We take on data where general-purpose AI fails. We build the evaluation before the model, hold our own work to it, and share measurements in conversations rather than on landing pages.

Repertoire outside the training distribution

The data: A Turkish music catalogue: microtonal modes (makam), asymmetric rhythmic cycles (usül, aksak) — repertoire general-purpose models were not trained on.
The constraint: Audio never leaves controlled infrastructure; no third-party services.
What we built: An end-to-end analysis pipeline — tagging, segmentation, lyrics transcription, aksak-meter detection, embedding similarity — with per-feature provenance and a curated makam/usül ground-truth set.
Status: Active. The analysis layer is working and evaluated against expert ground truth; the attribution layer on top of it is current research. We claim no more than the evaluation supports.

Outcome prediction from small clinical cohorts

The data: Gait and IMU sensor data; open clinical research corpora.
The constraint: Patient data stays off third-party clouds; small cohorts.
What we built: Prototype recovery-trajectory prediction and progress analytics, built with an industry partner.
Status: Concluded. Built, demonstrated and evaluated; the capability and the small-data evaluation discipline are retained.

The next case

Current work is under confidentiality — same shape: data that cannot leave the building, material outside the AI mainstream.

← How we work Working with us →