source: hugging face blog: building pakistan notice helper: a small ai tool for a very local safety problem
level: technical
pakistan notice helper is a safety-focused ai tool built for the hugging face build small hackathon. it helps people in pakistan understand suspicious messages before they click a link, call a number, share an otp, or make a payment. the tool accepts text or a screenshot and returns a risk label, a short explanation, visible red flags, and safe next steps. it does not claim a message is genuine or fraudulent. instead, it works as a triage tool to guide users toward safer actions.
the app supports both english and urdu. when a user switches to urdu, the interface changes to right-to-left layout and translates all headings, labels, and result controls. the model also generates the full safety response in clear urdu script. this makes advice easier to trust and act on. the tool looks for warning signs like urgent threats, requests for personal data, suspicious links or phone numbers, and impersonation of banks, couriers, or government departments. it then suggests verifying through independently found official channels.
the final model choice was qwen3.5 4b q8, served via llama.cpp on modal. earlier tests with larger models like qwen3.6 27b gave better quality but were too expensive and slow for a small, focused tool. a smaller vision-language model failed on reliability and accuracy. the 4b model balanced quality, speed, cost, and cold-start behavior. careful prompting and output contracts prevented the model from inventing urls or facts. the project also includes a privacy-safe public trace feature that records only limited metadata, not raw user content.
why it matters: it shows how a small, scoped language model can provide practical safety guidance for local problems without needing large, costly infrastructure.
source: hugging face blog: building pakistan notice helper: a small ai tool for a very local safety problem