Product direction
Make clinical speech easier to document, never harder to verify.
AfriHealth AI is a clinician-reviewed voice documentation prototype for community health teams working in Amharic-English. Intron Sahara v2.5 is the primary speech engine; clinicians remain responsible for reviewing transcripts and clinical outputs.
Prototype: functional
Production: not ready for identifiable data
1. Product modules
Module 1: Triage
Captures reviewed audio, returns code-switched transcription, and presents a clinician-review triage summary.
Module 2: EHR Intake
Supports draft documentation workflows. SOAP processing accepts structured input or falls back to manual review; prose-to-SOAP/ICD-10 generation is not complete.
Module 3: Post-Care
Supports follow-up simulations and highlights possible escalation signals for human assessment.
Active language
Amharic-English v1.0. Afaan Oromoo-English is Phase 2; Tigrinya-English is Phase 3.
Human review
Transcripts and generated clinical artifacts are drafts for qualified clinician verification.
Current boundary
Follow-up and triage support decisions; they do not diagnose, contact patients autonomously, or dispatch care.
Evidence snapshot
34.91%Normalized WER
57.78%Annotated target-term recall
42.22%M-WER
This is a saved-transcript comparison, not fresh model inference or population-performance evidence. Gemini reports 11.46% normalized WER; Whisper Tiny 99.56%; English-only Wav2Vec2 108.23%. A separate 15-recording clinical review found 56.38% mean WER, 44.33% target-term recall, and critical-term misses in 6 cases. All results require clinician review.
Review benchmark methodology and limitations
Readiness gates
Not ready for identifiable patient data: transcript retention/deletion controls, built-in clinician authentication and role-based access, and verified production gateway configuration remain open. Medication safety is not enforced end to end; medication details require independent clinician verification.
Provider credentials remain server-side, and private validation inputs stay outside Git. Complete clinician sign-off and backend safety controls before operational use.
Delivery priorities
01 · Close production safety gates
Implement transcript retention/deletion and clinician authentication with role-based access; verify gateway authentication and exact production CORS origins before identifiable data use.
02 · Complete clinician-reviewed workflows
Validate transcript-to-SOAP and coding behavior, require sign-off, and implement a backend medication safety gate before presenting medication checks as enforced.
03 · Strengthen evaluation evidence
Complete missing annotator metadata, retain model/configuration details, and expand consented, de-identified Amharic-English evaluation while keeping saved transcripts distinct from fresh inference.
04 · Verify release and handoff
Confirm frontend/backend deployment and CORS together, refresh the demo against the current build, and complete the recording and team handoff checklist.
Team and ownership
Project Manager · Architecture Lead
Ermias Amare
Clinical team
Hiwot Shiwangezaw
Rahel Tamiru
AI & ML research
Melaku Bayu
Software Engineer
Fasil Bazazew