Skip to main content

RescueBridge AI

RescueBridge AI is an offline-first, human-supervised prototype for collecting
multilingual disaster reports and turning them into reviewable incident records.
The current development path uses the explicit hosted Gemma API evaluation
adapter. Snowflake/CoCo is intentionally out of scope for this submission
build.

This project is a decision-support prototype. AI output is unverified and must
not be used for emergency dispatch or structural safety decisions.

Current status

See IMPLEMENTATION_STATUS.md. The repository is
being built in gates, starting with the local offline path.

Local prerequisites

  • Ollama with gemma4:e4b pulled locally
  • Python 3.12 or 3.13 recommended
  • Node.js and npm

Copy .env.example to .env and never commit .env. The current development
configuration selects the explicit hosted API mode; use AI_MODE=local when
you are ready to run the challenge-compliant offline path.

Hosted Gemma API mode

For the current API-based development path, copy .env.example to .env, set
GEMINI_API_KEY to a newly generated Google AI Studio key, and keep:

AI_MODE=hosted_gemma
ALLOW_HOSTED_EVALUATION=true
HOSTED_MODEL=gemma-4-26b-a4b-it

Install the backend dependencies, then run the API with:

python3 -m pip install -e './backend[dev]'
cd backend && python3 -m uvicorn app.main:app --host 127.0.0.1 --port 8000

The hosted evaluation endpoint is POST /api/v1/evaluation/gemini/analyze.
It requires is_synthetic: true and never silently falls back to a cloud
provider from the local path. The Google Gemma API model and image input flow
are documented by Google at
https://ai.google.dev/gemma/docs/core/gemma_on_gemini_api.

Do not send real survivor reports, personal data, or real emergency images to
the hosted endpoint.

Deliberately skipped challenge item

Snowflake/CoCo integration is not included by operator decision. The project
must not claim Snowflake evidence or mark that challenge item as complete.

License

Application source is MIT licensed. Gemma model weights and external datasets
retain their own licenses and terms.

Built with

geminiollamajavascriptnode.jspython

AI tools

Coded withGemma 4 26B A4B

Prize Categories Entered

  • Deepmind Google

    Best Use of Gemma 4

    Submitted
  • DigitalOcean

    Best Open-Source AI Project

    Submitted