Pakistan Notice Helper: Local AI Against Scams

An AI Tool for Digital Security in Pakistan

As part of the Hugging Face Build Small Hackathon, an innovative project has emerged: the Pakistan Notice Helper. This artificial intelligence tool has been designed to assist users in Pakistan in identifying and understanding the suspicious messages they receive daily. These messages, often disguised as official communications from banks, utility services, or authorities, can be genuine scams. The idea arose from the need to provide a simple and effective way to verify the authenticity of these messages before taking potentially risky actions.

Features and Language

The Pakistan Notice Helper stands out for its ability to operate in English and Urdu, thus meeting local linguistic needs. In Urdu mode, the application adjusts its interface for right-to-left display and fully translates content, including security assessments. This allows users to submit a suspicious message and receive a comprehensive security response in their chosen language, making the advice more accessible and applicable.

The application analyzes messages to detect warning signs such as urgent threats, requests for sensitive data like OTPs or card details, and suspicious payment links. It then provides security advice, such as verifying information through official channels rather than relying on links or phone numbers provided in the suspicious message.

Development and Technological Choices

The development of this tool highlighted the importance of selecting the right AI model. The project initially started with the Qwen3.6 27B model, which offered excellent quality, but deployment costs and practicality posed challenges due to the need for significant hardware resources. A trial with the MiniCPM-V 4.6 Q8 model was conducted in hopes of reducing costs, but it proved to be very slow on GPU, making its use impractical.

After several trials, the Qwen3.5 4B model was chosen for its efficiency and speed, while adhering to cost and resource constraints. This model allowed for a balance between quality and performance, which is essential for detecting security risks. The prompt and output contracts also played a crucial role, prohibiting URLs, phone numbers, organizations, and fabricated facts to ensure the reliability of the results.

Collaboration and Infrastructure

The project benefited from the support of Codex, which facilitated the rapid development of the custom frontend and backend. Codex was used as an engineering collaborator, allowing for inspection of the existing repository, implementation of modifications, testing, debugging issues, and maintaining alignment between the deployed system and the configurations of Modal, Gradio, and llama.cpp. Integration with Hugging Face Spaces via the Gradio server enabled the creation of a smooth user interface tailored to local needs.

The Urdu interface required additional work, as direct translations sounded unnatural, and certain layouts needed adjustments. This attention to detail transformed the tool into an effective and localized digital security solution, addressing the specific needs of users in Pakistan.

Lessons Learned

The development of the Pakistan Notice Helper provided several key insights. First, it demonstrated that smaller models can be highly effective when their scope is clearly defined. The model did not need to be a general investigator of scams but should focus on identifying visible risk signals and providing safe steps to follow.

Next, it was found that starting with a larger model, such as the Qwen3.6 27B, could offer exceptional quality but at the cost of complexity and high expenses. In contrast, the Qwen3.5 4B model provided the best balance among these factors, proving that a larger model is not always the best choice for a given product.

Finally, the importance of prompt and output contracts was emphasized, as they helped avoid common errors and improve the reliability of the system. The experience also showed that the user interface, particularly in Urdu, required special attention to ensure an optimal user experience.