00 · Decision support · Health & Sciences · Apache 2.0

Offline WHO IMCI paediatric triage. On the device.

Community health workers in low-connectivity settings type the symptoms. Gemma 4 on-device returns the IMCI tier and a structured referral pathway. A non-bypassable safety layer catches general danger signs even when the model is wrong. Nothing leaves the device.

Mustapha Alex · Medical intern, Nigeria Submitted to the Gemma 4 Good Hackathon May 2026

Apache 2.0 Gemma 4 · on-device Phase 1 Gate · 4 / 4

IMCI scenarios

4/4 PASS

Network packets

0non-localhost

Safety tests

17passing

Latency · M-series

9.8s median

Try a real triage→ Read the source→

Case · severe pneumonia

Airplane mode · 0 packets

PATIENT 11-month-old boy.
Cough for 3 days. Breathing 58/min.
Chest is sucking in. Restless and refusing to drink.

TierPINK

Refer urgently. Give first dose of injectable ampicillin + gentamicin (or oral amoxicillin if no injectable available). Transport to the nearest hospital immediately.

section Section 2 · severe pneumonia / very severe disease
danger_signs { chest_indrawing: true, refusing_to_drink: true }
confidence 1.00 safety_overrides none

01 · Evidence

Four canonical IMCI scenarios. Wi-Fi off. tcpdump running.

Scenarios drawn from the WHO IMCI Chart Booklet 2014. The product was gated on this verifiable airplane-mode test before any polish work.

eval/airplane-test-log.md

#	Scenario	Expected	Actual	Latency	Result
S1	Severe pneumonia · chest indrawing + refusing to drink	PINK	PINK	10.6 s	PASS
S2	Some dehydration · diarrhoea + restless + sunken eyes + slow skin pinch	YELLOW	YELLOW	10.7 s	PASS
S3	Uncomplicated malaria · RDT-positive fever, no danger signs	YELLOW	YELLOW	9.3 s	PASS
S4	Cough or cold · cough + drinking well + no chest indrawing	GREEN	GREEN	8.4 s	PASS

tcpdump$ wc -l /tmp/airplane-net.log → 0 // zero non-localhost packets across the entire run

02 · Architecture

One narrow path. No server. No analytics.

No fallback to anyone else's cloud. The same JSON output contract is designed for laptop V1 (Ollama, shipped) and Android V2 (LiteRT, roadmap).

infer.py · backends.py · safety.py

01 · Input

CHW symptom text (+ optional photo)

In-process. Never persisted to disk; never sent to any network address.

▼

02 · Model

Ollama @ 127.0.0.1:11434 → gemma4:e2b · 5.1 B · Q4_K_M

Native gemma4 architecture in Ollama ≥ 0.23.1. Avoids the iSWA-attention bug present in older llama-cpp GGUF distributions.

▼

03 · Parse

_extract_first_json · _validate_shape

Tolerant JSON coercion with a deterministic fallback so malformed model output cannot crash the UI.

▼

04 · Safety

safety.apply_safety_layer

R13 danger-sign keyword → force PINK · R14 confidence < 0.4 → append escalation · R15 adult case → refusal · R16 disclaimer banner.

▼

05 · Render

Gradio · tier card + pathway + reasoning + flags

On laptop, served on 127.0.0.1:7860. On Android (roadmap), the same JSON contract renders in a Kotlin native layout via LiteRT.

03 · Safety

Four invariants. Enforced after the model speaks.

17 unit tests cover the safety layer (laptop/test_safety.py). PocketTriage treats the classifier as fallible and never lets it have the final word.

laptop/safety.py

R13 · Danger-sign force-PINK

Override the model when general danger signs appear.

Keywords: unable to drink, vomits everything, convulsions, lethargic, unconscious, stiff neck, chest indrawing. Negation-aware: "no chest indrawing" does not fire. The tier is forced to PINK even if the model said GREEN.

R14 · Confidence floor

Escalate every low-confidence case to a human.

If the model self-reports confidence < 0.4, "Escalate to medical officer" is appended to the pathway regardless of tier.

R15 · Adult refusal

Adult conditions are out of scope — full stop.

Adult-specific terms (chest pain in adult, stroke in adult, pregnancy emergency) trigger a refusal with a referral to adult emergency protocol. Paediatric scope, hard boundary.

R16 · Non-dismissable disclaimer

The clinical-judgment ribbon is part of the layout.

"Decision support based on WHO IMCI 2014. Does NOT replace clinical judgment. Paediatric only (2 months – 5 years)." Always visible. Never closeable.

04 · Distribution

Outreach drafted with named contacts.

Three letters with full email bodies, shorter LinkedIn / WhatsApp variants, and send-log tables. Sending now that the submission is in and the demo URL is live.

outreach/*.md

Geneva

WHO Digital Health & Innovation

Dr. Alain Labrique · Director, DHI. Backup: Bernardo Mariano Jr.

Letter drafted

New Delhi · Mumbai

India NHM · ASHA programme

Smt. Roli Singh, IAS · Add. Secretary & Mission Director. + NHSRC + Maharashtra state Mission Director.

Letter drafted

Abuja · Awka

Nigerian NPHCDA

Dr. Muyi Aina · Executive Director NPHCDA. + Anambra SPHCDA + FMoH Coordinating Minister Pate.

Letter drafted

05 · Run locally

Clean clone to first triage — about fifteen minutes.

The same code path you see on the live Space, but on your own hardware. No network during inference. On Apple Silicon expect 8–11 seconds per triage.

README.md

# 1 · Install Ollama (macOS shown · Linux: install.sh from ollama.com)
brew install ollama
ollama serve &

# 2 · Pull the model (one-time, ~7 GB)
ollama pull gemma4:e2b

# 3 · Clone & install
git clone https://github.com/yonkoo11/pockettriage
cd pockettriage/laptop
python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 4 · Phase 1 Gate · same 4/4 you saw in the evidence table above
python eval_runner.py     # → 4 / 4 PASS

# 5 · Launch
python app.py             # → http://127.0.0.1:7860

Decision support based on WHO IMCI 2014. Does NOT replace clinical judgment. Paediatric only (2 months – 5 years). Photo input is treated as a supplementary signal; the pathway must always be groundable in the text description alone.