VoiceBridge — Offline Multilingual Clinical Triage AI

Fine-tuned Gemma 4 E4B for SATS 2023 / WHO ETAT-aligned clinical triage in resource-limited settings. No internet required. Runs fully offline via llama.cpp on hardware as small as a Raspberry Pi 5 8GB.

Submitted to the Gemma 4 Good Hackathon 2026 — Digital Equity Track.

Live demo: https://voicebridge.octo.net/ui/

Performance

Benchmarked on 100 real SATS-aligned clinical cases across 5 languages. Nothing simulated.

Metric Base Gemma 4 E4B VoiceBridge Fine-tuned
Exact match accuracy 85% 96%
Safe escalation rate 89% 100%
Unsafe under-triage cases 11 0
SATS validator agreement 76% 85%

Base Gemma 4 E4B uses the same custom prompt as my tuned model, which boosts its scores.

Per-level accuracy (fine-tuned): RED 100%, ORANGE 85%, YELLOW 95%, GREEN 100%, BLUE 100%

Per-language accuracy (fine-tuned): Bengali 100%, English 100%, Tagalog 100%, Hausa 95%, Swahili 85.7%

Zero unsafe under-triage cases. Every miss is a safe over-triage.

Languages

Benchmark languages: English, Swahili, Hausa, Bengali, Tagalog. Supports any language Gemma 4 speaks — Arabic, Turkish, Italian, Amharic, Hindi, French, and 140+ more without retraining. Multilingual capability comes from Gemma 4's native pretraining. Fine-tuning teaches clinical triage structure on top of that.

Intended Use

Community health workers, field nurses, and first responders in resource-limited settings — rural clinics, conflict zones, disaster response camps — anywhere without reliable internet access.

Thinking mode was disabled based on feedback from Medic's engineering team: speed and battery life matter more than reasoning depth on edge devices.

Usage

llama-cli -m voicebridge-finetuned-q4km.gguf \
  -p "<start_of_turn>system
You are a clinical triage assistant (SATS 2023 / WHO ETAT). Language: English.
Output ONLY a JSON object: triage_level (red/orange/yellow/green/blue),
primary_complaint, red_flag_indicators, recommended_action, confidence_score.<end_of_turn>
<start_of_turn>user
Patient not breathing, no pulse. CPR in progress.<end_of_turn>
<start_of_turn>model
{" \
  -n 600 --temp 1.0 --top-k 64 --top-p 0.95 -ngl 99

Training

  • Base model: google/gemma-4-e4b-it
  • Method: QLoRA via Unsloth, all 7 projection layers (q, k, v, o, gate, up, down)
  • LoRA rank 32, alpha 64, dropout 0.075
  • Dataset: 500 examples curated from SATS 2023 / WHO ETAT clinical scenarios
  • Languages: 8 language contexts, balanced across 5 triage levels
  • Hardware: RTX 5090, 2 epochs, AdamW 8-bit, bfloat16

Real-world Validation

Medic (medic.org), stewards of the Community Health Toolkit deployed to 100,000+ health workers across 15+ countries facilitating 8.7 million caring activities per quarter, reached out after seeing VoiceBridge. Their team expressed interest in an Android APK connecting to CHT via VoiceBridge's structured JSON output and asked us to prioritise Swahili and Bengali.

Links

Downloads last month
335
GGUF
Model size
8B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support