ReliefCopilot

From chaos to clarity: ReliefCopilot turns field notes into life-saving action plans, offline and 20× faster

Comment

Inspiration

In every disaster, the first few hours are chaotic responders take messy notes about water, shelter, lighting, or vulnerable people, but structuring those notes into a real action plan takes too long. We asked: what if AI could instantly turn chaos into clarity, completely offline, even when the internet is down? That’s how ReliefCopilot was born an AI assistant to help responders move faster, stay grounded in humanitarian standards, and save lives.

What it does

ReliefCopilot takes free-form, messy field notes and transforms them into:

-Structured Action Plans with clear tasks, priorities, owners, and timeboxes. -ICS-201 Style Briefings ready for rapid dissemination to teams. -Multilingual Communication messages (SMS/PA) in English, Hindi, and Telugu. -Evidence Citations directly from humanitarian standards (Sphere, WHO, FEMA, IFRC, ICS). -Safety Guardrails that block unsafe outputs like clinical dosing.

And it does all this offline, powered by open-source GPT-OSS models running on Ollama.

How we built it

Backend: FastAPI service that orchestrates Ollama’s GPT-OSS model with retrieval-augmented generation (BM25 index over humanitarian handbooks).

Corpus: Sphere, WHO, FEMA, IFRC, and ICS guidance ingested into a searchable local index.

UI/UX: A progressive web app with light/dark mode, JSON visualizations, and accessible offline-first design.

Model: gpt-oss-20B as the main engine (with quantized fallbacks like llama3.1-8B-q4 for lighter devices).

Dockerized: End-to-end setup for one-command deployment, fully offline, anywhere in the world.

Challenges we ran into

Running large models like 20B offline on typical laptops was tough — we had to optimize for Docker distribution and add quantized fallbacks.

Standard compliance: ensuring the AI doesn’t hallucinate but instead cites Sphere/WHO/FEMA text required careful corpus design and normalization.

UI/UX balance: making JSON action plans readable for responders under stress was harder than expected.

Safety: we had to hard-code lints to block dangerous medical dosage instructions.

Accomplishments that we're proud of

We built an end-to-end offline AI system that works with zero internet access.

Action plans are generated 10–20× faster than manual drafting.

Integrated multilingual support (English, Hindi, Telugu) for real-world field communication.

Successfully demonstrated retrieval-augmented generation with evidence citations from humanitarian standards.

Made the tool accessible and distributable via Docker so anyone can spin it up globally.

What we learned

AI alone isn’t enough grounding in trusted humanitarian standards is key for real-world trust.

Offline AI is possible: with careful optimization, even big models can run without internet.

Human-centered design matters: responders don’t want “AI essays,” they want clear tables, tasks, and briefings.

Hackathons are the perfect ground to mix AI research, UX design, and humanitarian impact into one project.

What's next for ReliefCopilot

Expand language coverage (e.g., Swahili, Arabic, Spanish) for wider global use.

Add voice input and TTS for responders who can’t type in the field.

Optimize for smaller, faster models while keeping accuracy and citations.

Partner with NGOs and disaster relief orgs to test in simulation drills.

Explore fine-tuning GPT-OSS on humanitarian datasets for even more accurate, context-aware action plans.

Built With

Share this project:

Updates