Abid Ali Awan developed the Pakistan Notice Helper AI tool for the Hugging Face Build Small Hackathon. The application accepts text or screenshot inputs and generates a risk label, a short explanation, visible red flags, and safe next steps.
The tool does not verify whether messages are officially genuine or fraudulent. It scans for indicators including threats of account suspension and requests for one-time passwords (OTPs), personal identification numbers (PINs), passwords, card verification values (CVVs), and Computerized National Identity Card (CNIC) data. The application supports English and Urdu interfaces, including Roman Urdu text input. The Urdu interface uses a right-to-left layout, translates all interface elements, and generates safety assessments in Urdu script.
The deployed model is Qwen3.5 4B Q8 MTP GGUF with a vision projector. The Build Small Hackathon restricted model sizes to a maximum of 32 billion parameters. Awan initially tested the Qwen3.6 27B model during development; however, this model required high VRAM capacity and experienced extended cold-start durations. The developer selected the Qwen3.5 4B model for production to reduce hosting costs and improve startup speed.
The tool advises users to verify information through contact details found via independent, official channels rather than using links or numbers provided in suspicious messages. Awan disabled the model's internal reasoning feature to prevent the 500-token output budget from being exhausted before JSON generation. The token allocation for image processing requests was increased after a screenshot input reached the initial completion limit. The system prompt was updated to prohibit the generation of unverified URLs, phone numbers, organizational names, and unconfirmed facts. Awan, a software developer, stated, "Building with small models is less about chasing the highest benchmark score and more about finding the right balance between quality, speed, cost, and product safety." He added, "Pakistan Notice Helper does not need to be a general scam investigator."
A test deployment of the MiniCPM-V 4.6 Q8 model encountered quota limits and execution errors on the ZeroGPU platform. Interface testing determined that a bundled Nastaliq webfont decreased readability in the mobile layout. The system architecture connects a Hugging Face Space frontend to a queued Gradio Server endpoint, a Modal backend, and CUDA llama.cpp.
No independent assessment of Pakistan Notice Helper’s claims was available.