Building Pakistan Notice Helper: A Small AI Tool for a Very Local Safety Problem

Source: HuggingFace Date: 2026-06-08 URL: https://huggingface.co/blog/build-small-hackathon/building-pakistan-notice-helper

Summary

A hackathon writeup documenting the build of a scam-detection tool for users in Pakistan, analyzing suspicious messages from entities impersonating banks, couriers, and government agencies. The author settled on a quantized Qwen 4B model over a 27B variant — scoring ~80 vs ~95 on task quality — deployed via llama.cpp on Modal with a Gradio frontend on Hugging Face Spaces, and added full Urdu (RTL) support. Regression testing reached 100% pass rate across 10 cases.

Implications

• Small-model viability: The project is a clean data point that a 4B quantized model can reach production-quality accuracy (80/100) on a well-scoped, single-task problem — cost and latency advantages compound when the task surface is narrow. Feeds the model-landscape thread on capability-per-dollar at the small end.
• Localization as product work: Urdu RTL support required non-trivial engineering beyond translation; signals that agentic/AI tools targeting non-English markets carry UX debt that benchmark scores don’t capture.
• Open agentic tooling stack: The Modal + llama.cpp + Hugging Face Spaces architecture is a repeating pattern for low-cost inference deployments — relevant to the dev-tooling thread on inference infra democratization.
• Trust and safety framing: A safety tool that explicitly avoids claiming definitive authenticity (triage, not verdict) is a useful design precedent for high-stakes consumer AI — relevant to trust/supply-chain thread.