Echo Reach: Last-Mile Climate Voice Alerts

Project Overview

Echo Reach is an automated climate early warning platform engineered to protect vulnerable, off-grid communities. By monitoring meteorological telemetry and utilizing generative AI, the system bypasses internet and smartphone dependencies. It translates complex climate data into local languages and delivers actionable voice advisories directly to standard feature phones.

This project was developed for the IGAD Hackathon 2026 to address climate communication barriers ahead of the 74th Greater Horn of Africa Climate Outlook Forum (GHACOF 74) scheduled for August 17 to 18, 2026, in Kigali, Rwanda.

System Architecture

The platform operates on a strict five-stage execution pipeline:

1. Telemetry Polling: A Node.js core continuously samples Open-Meteo API endpoints for extreme precipitation and temperature anomalies.
2. AI Simplification: Raw meteorological bulletins are processed through Google's Gemini API to extract hazards and generate concise, actionable instructions. The pipeline implements a multi-tier model fallback strategy (gemini-3.6-flash to gemini-3.5-flash-lite) to manage API quotas.
3. Neural Translation: The simplified text is translated into regional languages utilizing Meta's NLLB-200 model running locally via Transformers.js.
4. Speech Synthesis: A dedicated Python FastAPI microservice converts the translated text into natural audio utilizing Meta's MMS-TTS models.
5. Telephony Dispatch: The system triggers the Africa's Talking API, executing outbound voice callbacks that stream the generated .wav audio over the cellular network, accompanied by a redundant SMS backup.

Technology Stack

• Core Backend: Node.js, Express.js
• Database: MariaDB with Sequelize ORM
• Artificial Intelligence: Google Gemini SDK (@google/genai), Hugging Face Transformers.js, Meta MMS-TTS
• Telephony Gateway: Africa's Talking Voice and SMS APIs
• Production Infrastructure: Ubuntu Linux VPS, PM2 Process Manager, Nginx Reverse Proxy

Engineering Resilience

• Hardware Optimization: The system executes a 1.2GB NLLB-200 translation model alongside Python speech synthesis on a heavily constrained 2GB virtual private server. Memory stability is maintained by utilizing 8-bit ONNX quantization, swap memory expansion, and a strict singleton pattern to prevent Linux kernel Out-Of-Memory (OOM) process terminations.
• USSD-Ready Architecture: The backend user onboarding flow is structured to support USSD menu integration, ensuring that target users can subscribe to regional alerts entirely offline.

License

MIT License