PaperSight

Clinical triage assistant that photographs paper intake forms, extracts patient data with a vision LLM, and assigns priority scores using deterministic clinical rules — fully offline on a Mac Mini M2 with 8GB RAM.

Powered by Gemma 4 E2B via Ollama for on-device serving.

Gemma is a trademark of Google LLC.

Built for the Kaggle "Gemma 4 Good Hackathon" (deadline: May 18, 2026).

---

What it does

1. Photograph a paper patient intake form with any smartphone, tablet, or desktop browser
2. Extract fields (name, age, gender, chief complaint, duration, allergies, medications, referred by) using Gemma 4 E2B via Ollama
3. Triage the patient with rule-based scoring derived from WHO ETAT + SATS + MSF guidelines
4. Queue patients by priority (Red → Yellow → Green) with filtering and detail views

Only Age and Chief Complaint are required for triage. All other fields are optional and collected at bedside if unreadable.

---

Architecture

---

Requirements

• macOS (tested on Mac Mini M2, 8GB RAM)
• Ollama v0.21.1+
• uv for Python package management

---

Quick Start

1. Install dependencies

uv sync

2. Pull the model (one-time, ~7.2GB)

ollama pull gemma4:e2b

3. Pre-warm Ollama (avoids ~57s cold-start)

ollama run gemma4:e2b "say ready"

4. Start the server

PYTHONPATH=src uv run uvicorn papersight.main:app --host 0.0.0.0 --port 8000

Note: PYTHONPATH=src is currently required due to a module resolution issue with the papersight-server console script. Use 0.0.0.0 (not 127.0.0.1) so other devices on the same Wi-Fi network can reach the server.

5. Open in browser

• On this Mac: http://127.0.0.1:8000
• On phone / tablet (same Wi-Fi): find your Mac's local IP with ipconfig getifaddr en0, then open http://<YOUR_IP>:8000

---

API Endpoints

Method	Endpoint	Description
GET	/	Serve frontend
POST	/api/extract	Upload photo → extract + triage
GET	/api/patients	List all patients (priority sorted)
DELETE	/api/patients/{id}	Remove a patient

---

Caveats & Limitations

PaperSight is clinical decision-support, not a replacement for trained clinical judgment. A nurse or doctor must always verify the triage and examine the patient.

• Deterministic triage — rule-based (WHO ETAT + SATS + MSF), not LLM-based. Auditable and reproducible, but cannot reason about novel or atypical presentations.
• Brittle extraction — vision model output depends on handwriting, blur, and lighting. Hyphenated compounds are normalized (chest-pain → chest pain), but there is no stemming or synonym expansion. Unreadable fields are marked "unreadable" and collected at bedside.
• No vital signs — triage is symptom-based only. BP, SpO₂, pulse, and temperature are not captured.
• Offline constraints — no cloud means no model updates, no sync, and no remote backup. The Mac Mini is a single point of failure.
• Hardware ceiling — 8GB RAM target. Image uploads capped at 10 MB. Context window limited to 4,096 tokens. No batch processing.

---

License

Apache 2.0