Cloud inference costs for continuous background tasks (like news summarization) stack up quickly. I set out to answer a question: Can an old, unused smartphone serve as a dedicated, zero-cost AI server for real production traffic?
I repurposed a 7-year-old Android phone as a self-hosted AI inference server running Qwen 1.5B via llama.cpp inside Termux.
Pipeline Architecture
- Self-Hosted Inference Endpoint: Configured Termux with C++ build tools, compiled
llama.cppfor ARM64 architecture, and hosted Qwen 1.5B with a local REST API endpoint accessible on the local network. - Automated Scraping & Processing: Built a Python backend that scrapes RSS feeds across major news sources every 2 hours, extracts article content, and sends summarization prompts to the phone server.
- TikTok-Style Scroll Feed: Summaries are delivered to a modern vertical scroll web app at
everythingblog.in, reaching over 1,000+ daily active users at peak — with zero cloud compute or API costs.
Key Technical Achievements
- Transformed a 7-year-old Android device into a 24/7 ARM64 AI inference server.
- Automated end-to-end RSS extraction and LLM summarization pipeline.
- Scaled feed to 1,000+ daily active users with zero server or API costs.