Building Pakistan Notice Helper: A Small AI Tool for a Very Local Safety Problem
Pakistan Notice Helper is a small AI safety tool for Pakistan that flags risky messages, explains warning signs, and suggests safer next steps.
Intelligence analysis by GPT-5.4 Mini
Built for a hackathon, Pakistan Notice Helper uses a small Qwen3.5 4B model to triage suspicious SMS, notices, and screenshots in English or Urdu. It focuses on red flags and safe next steps, not on proving whether a message is truly genuine.
It is like a warning flashlight for suspicious messages. Instead of proving if a note is real, it points out danger signs and tells the person to check safely through official channels.
Analysis
What the app is
Pakistan Notice Helper is a safety-focused AI tool built for people in Pakistan who receive suspicious messages that may claim to come from banks, couriers, tax offices, traffic police, utilities, mobile operators, or government departments. The system accepts either text or a screenshot and returns a risk label, a short explanation, visible red flags, and safer next steps.
The project is intentionally not an authenticity checker. It does not try to prove that a notice is definitely real or definitely fake. Instead, it is designed to help a person pause, spot warning signs, and avoid risky actions like clicking a link, calling an unknown number, sharing an OTP, or making a payment.
Model and product choices
The builder started with larger and smaller models before landing on Qwen3.5 4B Q8 through llama.cpp as the production choice. The article says the larger Qwen3.6 27B model produced excellent results, but it was too heavy, expensive, and slow for a hackathon demo with irregular traffic. A smaller MiniCPM-V 4.6 Q8 experiment was not reliable enough and struggled with quality.
The final model was chosen because it balanced quality, speed, cost, and cold-start behavior. The article also notes that the stack could handle both text and screenshots while staying under the hackathon’s 32B model limit.
Urdu, safety, and output discipline
A major product decision was full support for English and Urdu. In Urdu mode, the app switches to right-to-left layout, translates the interface, and asks the model to answer in clear Urdu script. That matters because suspicious messages in Pakistan are often written in English, Urdu, Roman Urdu, or a mix of all three.
The builder also tightened prompting and output contracts after early failures. One version used too much of the token budget, and another generated an official-looking domain that had not been verified. The fix was to disable thinking for production and forbid invented URLs, phone numbers, organizations, and facts.
Results and next step
The article says the final regression suite reached 10 of 10 passes, with all high-risk scam cases and both screenshot cases passing. The next planned feature is an agentic verification workflow that would search for scam warnings, compare claims against independently found sources, and help judge whether a notice is original, copied, or already reported as a scam.
Key points
- Pakistan Notice Helper is a local safety tool for suspicious messages in Pakistan.
- It supports both English and Urdu, including right-to-left UI and Urdu-script output.
- The production model became Qwen3.5 4B Q8 after larger and smaller options were tested.
- The builder tightened prompts and output rules after early failures with invented details.
- Public traces share only limited metadata, not raw screenshots or full user messages.
If the tool keeps improving, it could help people in Pakistan slow down before sharing OTPs, bank details, or money with scammers. The Urdu support and screenshot handling make it easier for more users to understand the warning quickly.
It still does not verify authenticity, so some risky messages could be missed or some real notices could be treated as suspicious. The article also says live text and images still go to a private inference endpoint, so users must trust the system’s handling of sensitive content.
