AMS-Gemma4: Offline Antibiotic Stewardship AI

Offline clinical AI for rural Indian hospitals. Gemma 4 + native function calling + Tamil output. WHO AWaRe 2023 + ICMR India data. No internet needed.

Comment

Inspiration

I am a pharmacy professor in Puducherry, India. Every day doctors in rural hospitals make antibiotic decisions without a pharmacist. Antibiotic resistance kills more people than HIV and malaria combined. India bears the highest burden globally. I built AMS-Gemma4 to give every rural doctor a clinical pharmacist in their pocket.

What it does

AMS-Gemma4 is an offline antibiotic stewardship AI. Doctor enters patient data and culture report. System calculates CrCl using Cockcroft-Gault, checks WHO AWaRe 2023, looks up ICMR India resistance data, applies safety gate, and generates 10-section recommendation in English or Tamil. Key result: Nitrofurantoin automatically REJECTED when CrCl < 60 — the error that harms elderly patients daily.

How we built it

Gemma 4 E2B via Ollama for offline inference. Python function calling layer for guaranteed clinical math. Gradio web interface. WHO AWaRe 2023 + ICMR India 2024 knowledge base. Multimodal culture report image upload. Tamil output via prompt engineering. Resilient fallback — when API fails, local Ollama takes over automatically.

Challenges we ran into

Ensuring mathematical accuracy in clinical decisions — solved with Python pre-computation before Gemma reasons. HuggingFace deployment with Gemini API quota limits. Tamil clinical terminology validation. Making it truly offline on 8GB MacBook M1.

Accomplishments that we're proud of

Works completely offline on 8GB RAM. First antibiotic stewardship AI with Tamil language output. Catches real clinical errors automatically. Validated against 4 real clinical scenarios. Apache 2.0 — free for any hospital to deploy today.

What we learned

Clinical AI requires guaranteed accuracy — not probabilistic guessing. Python function calling before LLM reasoning was the critical architectural decision. Offline deployment is not a limitation — it is a feature for rural healthcare.

What's next for AMS-Gemma4: Offline Antibiotic Stewardship AI

Deploy to 25,000+ primary health centres across India. Add Telugu, Kannada, Hindi language support. Partner with ICMR for real-time resistance data updates. Build Android mobile app for offline use. Expand to other clinical decision support areas.

Built With

Share this project:

Updates