Illustrative pilot projection
~5 minutes
Average consultation time saved
Planning assumption per patient; not measured in a clinical study.
Prototype impact view
Challenge alignment
Clinical Impact & Efficiency Dashboard
Illustrative workflow-value assumptions for discussion and pilot planning. These figures are not observed results, validated clinical outcomes, or findings from a clinical study.
Estimated Clinical Workflow Impact (Pilot Projections). Not validated clinical study findings.
Illustrative pilot projection
~200 minutes
Estimated daily time saved per clinician
Illustrates 5 minutes across an assumed 40 consultations per workday; neither input is a measured outcome.
Illustrative pilot projection
~70%
Documentation burden reduction
Hypothesis for future evaluation; not yet validated or measured.
Illustrative pilot projection
~15%
Scaled patient support
Projected daily capacity contribution; definition and real-world impact require prospective measurement.
Production-readiness & trust indicators
Prototype controls only; these badges are not production certification.
- ✅ Responsible AI Guardrails
- ✅ Clinician Approval Required
- ✅ FHIR Export Protected by Review Gate
- ✅ Benchmark-Driven Model Evaluation
Not ready for identifiable patient data: authentication/RBAC, transcript retention and deletion controls, auditability, and production deployment configuration still require implementation and verification. Benchmark findings are limited to their documented dataset and methods.
MODULE 1
Frontline Triage Support
Supporting community health workers (CHWs) with voice-driven Amharic-English triage documentation.
Noise Suppression: Active
Mode: Amharic-English Code-Switch
5 Illustrative Demo Fixtures: Select a Case or Record Live
Ready to record or run pre-loaded sample
Review saved recording before upload
Download local copy
Audio processed via Intron Sahara v2.5 ASR Engine
Intron Sahara v2.5 Code-Switched ASR Stream
Click a preset above or tap the microphone to begin code-switched speech recognition...
Extracted Clinical Reference Entities
Awaiting audio processing...
Triage Reference Category (Unverified)
Pending Processing
Clinical Reference Suggestion: Pending Verification.
MODULE 2
Voice Intake & EHR Integration Engine
Automating clinical note-taking during doctor consultations with structured FHIR / EHR JSON output.
Idle
Select the speaker before each segment.
LIVE CODE-SWITCHED CONSULTATION TRANSCRIPT
Click "Start Consultation Recording" to capture live audio, or "Load Consultation Demo" for a scripted example...
Record the live doctor-patient conversation; the chart auto-fills as fields are recognized.
Extracted fields are clinical reference suggestions and require clinician review before any record action.
Structured SOAP Clinical Reference Draft
Editable clinician-support draft. Streaming updates fill fields that have not been manually changed; all content requires clinician verification.
The backend validates structured SOAP JSON; prose returns a manual-review draft and does not generate ICD-10 codes.
Human review required. This clinical reference draft must be verified before any FHIR export or EHR commit.
Auto-Filled EHR Medical Chart
Clinician Verification Gate
Review Required before FHIR Commit
Awaiting consultation processing...
Review Required before FHIR Commit
FHIR JSON Preview
Workflow Simulation
VoiceBot Follow-Up Workflow Demo
Prototype demonstration only. This state machine does not schedule or place a real call, contact a patient, or submit an alert to clinical staff.
- Step 1 Patient Discharged / Intake Completed Current simulated state.
- Step 2 Post-Treatment Reminder Scheduled Pending in simulation.
- Step 3 VoiceBot Adherence Check Triggered Pending in simulation.
- Step 4 Follow-Up Status & Flagged Adverse Reactions Log Pending in simulation.
Workflow Simulation: Patient Discharged / Intake Completed.
Workflow Simulation — Example adverse-reaction flag:
Patient-reported rash after medication in this fictional scenario. Flagged for clinician review only; not clinically verified, not a diagnosis, and no clinical action has been taken.
MODULE 3
Post-Care Voice Follow-up
Clinician-support voice follow-up simulations with risk indicators for human review.
Post-care demo focus: illustrative fixture 2 of 5; separate from the clinical benchmark.
Voice Follow-up Demo
Outpatient Call Simulator
Patient:
Lemlem (Maternal Post-Op Follow-up)
Post-Op:
Day 3 (Appendectomy)
Language:
Amharic-English Code-Switch
Live Post-Care Protocol Analyzer
Applies the configured post-care protocol to supplied transcript text. Risk and acuity results are clinical reference suggestions and require clinician review.
Nurse review suggested
Decision support from reported symptoms only -- a nurse/doctor must confirm with real vitals and exam.
Real Voice Follow-up Call
Interactive voice session: Intron speaks each prompt, you answer by microphone, and the service transcribes the response. Returned questions and risk indicators are clinical reference suggestions for clinician review; this panel does not automatically contact clinical staff.
Call Idle
Select a scenario above to start outbound AI call
Outbound Voice Interaction Transcript
Press a simulation button on the left to initiate the voice call...
Parsed Status
No Active Call
Clinical reference indicators are derived from the supplied transcript and require clinician review.
Escalation Queue
Monitoring outpatient recovery telemetry...
Judge reference
Sahara CodeSwitch Africa Challenge Alignment
Requirement-to-evidence map for a fast review. Evidence claims below describe repository contents and prototype behavior; they are not a claim of clinical efficacy or production readiness.
| Requirement | Project Implementation | Evidence Location |
|---|---|---|
| Voice downstream clinical workflows | Browser audio capture and transcription, clinician-review triage and intake surfaces, FHIR-compatible export with review gate, and post-care API workflows. The added VoiceBot state machine is a separate local workflow simulation; it does not place or schedule calls. | Triage and intake modules; VoiceBot simulation; ARCHITECTURE.md; main.py routes `/api/v1/post-care/analyze`, `/api/v1/fhir/export`, `/api/v1/ehr/commit`. |
| Three-or-more speech model comparison | A six-adapter evaluation pipeline is wired to the 100-case master manifest. Current source transcripts are not independently audio-verified, and mock outputs are not accuracy evidence; no live comparison is represented as validated benchmark performance. | BENCHMARK_RESULTS.md; 100-case master manifest; inference_engine.py; evaluator.py. |
| Health category focus | Healthcare documentation and voice-support prototype focused on clinician-reviewed Amharic-English code-switching. SOAP/ICD output is not automatically generated from ordinary prose; clinical artifacts require human verification. | README.md; ARCHITECTURE.md; CLINICAL_RECORDING_PROTOCOL.md. |
| Responsible AI, safety, and human review | UI disclaimers and review badges; explicit client review confirmation before export; backend sign-off assertion required by FHIR/EHR endpoints. Reviewer identity/RBAC, auditability, retention/deletion, subgroup fairness, and clinical efficacy are not established by this prototype. | RESPONSIBLE_AI.md; tests/test_safety.py; index.html; main.py. |
| Working prototype and judge-ready demonstration | Single-page browser prototype with one-click impact dashboard, VoiceBot workflow simulation, challenge map, clinical modules, and benchmark view. Confirm the linked video matches this build before submission. | DEMO_SCRIPT.md; SUBMISSION_READINESS.md; README.md. |
Evidence note: the 100-case master manifest contains source text, not independently audio-verified gold transcripts. Mock outputs are synthetic and not model-performance evidence. Do not present the VoiceBot simulation or projected impact values as measured or live patient outcomes.
EVIDENCE REVIEW
ASR Benchmark and Clinical Validation
The matrix below uses five illustrative application fixtures, separate from the 100-case clinical benchmark. Its scores are synthetic examples and must not be used for model ranking or performance claims.
Fixture benchmark limitation
The displayed matrix uses embedded demonstration hypotheses and illustrative case values. Its WER/CER values are synthetic fixture examples, not independent audio inference or proof of production superiority. The five demo fixtures are separate from the authoritative CS-01 to CS-100 clinical manifest.
Clinical benchmark status
The active clinical corpus is benchmark/metadata/BENCHMARK_MANIFEST.csv (CS-01 to CS-100). Source text and provisional labels require audio-aligned human verification; mock outputs are synthetic, and validated performance metrics are not available until audio verification is complete.
100-Case Clinical Benchmark Pipeline
The CS-01 to CS-100 source text is provisional, mock outputs are synthetic, and performance claims require consented audio plus audio-aligned human verification. Mock runs report pipeline coverage only.
Verified model WER
N/A
Verified term recall
N/A
Clinical term error rate
N/A
| Baseline | Normalized WER | Target recall | M-WER |
|---|---|---|---|
| Gemini | N/A | N/A | N/A |
| Whisper | N/A | N/A | N/A |
| Wav2Vec2 | N/A | N/A | N/A |
No verified live model scores are currently available. Mock outputs are intentionally excluded from accuracy metrics. See BENCHMARK_RESULTS.md for the current evidence status and limitations.
Dataset: 5 illustrative UI fixtures
Language: Amharic-English
Demo Fixture JSON payload (GOLD-ETH-001)
Illustrative Demo Fixture WER vs CER (Not Measured Results)
Aggregated CorpusClinician Human-In-The-Loop (HITL) Scores
Factuality and safety are qualitative human-evaluation fields; no case-level ratings are supplied for this demo fixture.
| Speech AI Model | WER | CER | Factuality | Safety |
|---|
Clinician Review Notes:
No case-level clinician ratings or hallucination analysis are available for this fixture; displayed scores are transcript metrics only.
Live Intron Sahara v2.5 Benchmark
Upload or record an audio sample in Amharic-English to evaluate live Intron Sahara v2.5 ASR performance.
No recording
Target ASR Engine:
⚡ Intron Sahara v2.5
| Provider | Status | Latency | WER | CER | Transcript |
|---|
ℹ️ Levenshtein Diff is enabled for illustrative demo fixture references or when a reference transcript is provided in the calculator.
Live Levenshtein Edit-Distance Calculator
Formula: WER = (S + D + I) / N
[+ Inserted (Green)]
|
[- Omitted (Red Line-through)]
Calculated WER: --%
Calculated CER: --%