Salud Andina: testing Gemma 3n for rural health guidance in Ecuador
An inference-only Kaggle notebook testing Gemma 3n E4B on image and text questions in Spanish, as groundwork for a rural health guidance idea submitted to the Gemma 3n Impact Challenge.
Context and problem
People in rural Ecuador can be far from a clinic and have limited connectivity. The notebook asks a narrow question: can Gemma 3n, running on a Kaggle GPU (T4 x2), give basic health orientation in Spanish? The project idea is a rural health assistant that takes text and voice queries, gives preliminary orientation and flags potentially urgent situations. It is a guide, not a doctor: it never diagnoses and always points to a professional for serious symptoms.
Constraints
- Hardware: the notebook requires a Kaggle T4 x2 accelerator.
- The model runs 4-bit quantized with a maximum sequence length of 1,024 tokens.
- The notebook is inference only, with no training data.
Architecture
How to read it. Read left to right. A system prompt, a user question and a test image go to the model, and the response is judged by reading it. The highlighted box is the model: the notebook only tests how Gemma 3n behaves, it does not train or measure it.
The notebook loads Gemma 3n E4B instruction-tuned in 4-bit through Unsloth. A system prompt frames the model as a health assistant for rural South America: culturally sensitive, simple language, never diagnose, always recommend a doctor for serious symptoms. The prompt asks for responses in Spanish.
Decisions and trade-offs
- Inference only. I tested the model as it is, without fine-tuning or a training dataset.
- 4-bit E4B. The notebook loads one 4-bit variant of Gemma 3n and runs it on the Kaggle T4 x2 environment.
- A guardrail in the prompt. The system prompt states the never-diagnose rule and the referral rule before any question is asked.
Governance and cost
The three test images used in the notebook are not published here. The system prompt sets the boundary: no diagnosis and a doctor for serious symptoms.
Outcome
Assessment is qualitative and mine alone, with no accuracy, latency or benchmark numbers. In the tests the model answered in Spanish and offered recommendations on what to do. With an unrelated image, a toy bear, it still gave a coherent response and stayed in Spanish. I read that as a sign the model can handle multimodal questions in a basic assistance setting, not as evidence of clinical value. The video presents the project proposal for the Gemma 3n Impact Challenge; it is a proposal, not a record of what the notebook built.
What comes next
The planned solution has four parts: conversational AI where people describe symptoms in their own words, a knowledge base of local medicinal plants and common generic medicines, symptom evaluation with next steps, and a PDF summary to share with a doctor. None of these is built in the notebook.