Abhisek Roy, a developer participating in the Hacktoberfest Weekend Challenge, has released "Offline SMS Buddy," a lightweight application designed to help non-technical users identify potential scams in text messages. Built specifically for his mother, who frequently struggled to distinguish between legitimate bank notifications and fraudulent alerts, the tool runs entirely on a local laptop. The project demonstrates how open-weight models can solve real-world privacy concerns without relying on cloud APIs.

The Stack: Qwen 2.5 and Ollama

The application leverages Qwen 2.5, an open-weight large language model, specifically the 3B parameter version to ensure compatibility with standard hardware featuring 8 GB of RAM. Roy used Ollama to serve the model locally at localhost:11434, paired with Streamlit for the user interface and Python with uv for packaging. The core functionality asks the AI to answer three specific questions: what the SMS means, what action is required, and what risks are present. By using a Pydantic schema, the developer enforced structured JSON output, preventing the model from generating ambiguous free-text responses.

Safety Nets and UX Design

Recognizing that small models can hallucinate or miss critical warnings, Roy implemented a dual-layer safety system. While the AI provides detailed explanations, a separate regex-based check scans for keywords like "KYC," "blocked," or shortened URLs. If these terms appear, the app forces a red warning banner regardless of the AI's output, ensuring that critical scam indicators are never missed. The user interface was stripped down to a single text box and a large button, with errors displayed in plain language rather than stack traces, making it accessible for users unfamiliar with command-line tools.

Why Local Inference Matters Here

The decision to run the model offline was driven by strict privacy and reliability requirements. SMS messages often contain sensitive data such as partial account numbers, OTPs, and balances. Sending this data to a cloud provider introduces risks related to data retention and third-party logging. With Offline SMS Buddy, data never leaves the device. Additionally, the local setup eliminates recurring API costs and ensures functionality during internet outages, a crucial feature for a tool meant to provide immediate reassurance during urgent situations.

Key Takeaways

  • Privacy First: Local LLMs allow sensitive personal data, like bank SMS, to remain on-device, avoiding cloud exposure.
  • Hybrid Safety: Combining AI explanations with deterministic regex checks creates a more robust defense against scams than AI alone.
  • Hardware Accessibility: The Qwen 2.5 3B model runs effectively on consumer laptops with 8 GB RAM, lowering the barrier for local AI adoption.
  • Cost Efficiency: Zero per-request costs make the tool sustainable for long-term personal use without subscription fatigue.

The Bottom Line

Roy’s project is a masterclass in pragmatic AI engineering: it proves that you don’t need massive parameters or cloud subscriptions to solve a specific, high-stakes problem for real people. By prioritizing local execution and deterministic safety checks over pure LLM reliance, Offline SMS Buddy offers a blueprint for building trustworthy, private tools for the non-technical majority.