We're building Amharic clinical speech and note-generation models from the ground up. As we validate and publish, the work will live here — methods, accuracy, and privacy, in the open.
The research program behind Bekur's scribe — published as it's validated. Each paper documents the methods, datasets, and results so the work can be read, checked, and built on.
How we recognize the mixed Amharic and English used in real Ethiopian consultations, including medical terminology, accents, and noisy clinic environments.
Bekur Research
Turning a free-flowing consultation into an accurate, structured clinical note — the model design, prompts, and the guardrails that keep it faithful to what was actually said.
Bekur Research
How we evaluate fidelity, completeness, and hallucination rates against clinician-written notes, and the human-in-the-loop review that keeps the scribe safe to use.
Bekur Research
Our approach to protecting patient data — consent, de-identification, data residency, and the design choices that keep sensitive clinical conversations private.
Bekur Research
More coming soon. These papers are in active preparation — request early access to receive each one as it's published.
Off-the-shelf models weren't built for Amharic or for Ethiopian clinical reality. So we build — and study — our own. Four questions drive the work.
Speech and language models that understand Amharic, English, and the code-switching real consultations are full of — including medical terminology and local clinical context.
Rigorous evaluation against clinician-written notes, so the structured note the scribe produces is faithful, complete, and safe to rely on.
Consent, de-identification, and data-residency methods that protect sensitive patient conversations in low-resource healthcare settings.
How an AI scribe fits the day of an Ethiopian clinician — what to surface, when to defer to a human, and how to give time back to patient care.
The papers are in active preparation. Tell us you'd like early access and we'll send each one as it lands — methods, results, and all.