📰 Key Takeaways

Pakistan Notice Helper is a compact AI safety tool developed to address local scam message issues in Pakistan, built by a developer during the “Build Small” hackathon Backyard AI track. Pakistani users have long been receiving suspicious messages impersonating banks, courier companies, tax authorities, telecom operators, or government agencies. Identifying what’s fake isn’t the hard part—the real challenge is not knowing what to do before clicking links, making calls, providing OTPs, or making payments. This tool isn’t an “authenticity checker”—it’s a risk classification tool: users can input text or screenshots, and the system returns risk level labels, brief explanations, visible warning flags, and safe follow-up recommendations.

On the technical side, the developer initially tested larger Qwen models, ultimately settling on the Qwen3.5 4B Q8 quantized version, running via llama.cpp on CUDA, and connecting it to a Modal endpoint, Gradio Server, and a custom front-end built on Hugging Face Space. The overall model size is well under the hackathon’s 32B limit, while also supporting both text and screenshot bimodal processing. After evaluating ten test cases (including high-risk scams and screenshot scenarios), all passed.

Language support is a critical product decision: suspicious messages in Pakistan are often written in a mix of English, Urdu, or Roman Urdu. So the tool supports both languages—when switching to Urdu mode, the interface automatically adjusts to right-to-left layout, and the model generates a complete evaluation report in Urdu, including risk labels, explanations, warning flags, and suggested response drafts. The warning signals the tool detects include: account freeze threats, requests for OTP or CNIC identity information, suspicious payment links, and impersonation of financial institutions or government agencies.

💬 JudyAI Lab’s Perspective

What Pakistan Notice Helper solves isn’t the technical problem of “determining authenticity”—it’s the behavioral gap of “not knowing what to do next.” This is an angle we rarely see being directly addressed when looking at AI safety tools.

There are a few notable design decisions in this case. The developer chose not to use the 32B model limit and went with the Qwen3.5 4B Q8 quantized version—a pragmatic trade-off between performance and deployment cost. More importantly is the output design: instead of giving a binary “real/fake” verdict, the tool returns risk levels, warning flags, and actionable follow-up recommendations—so users know what to do next after receiving the result, not just confirming “it’s a scam.” Language support also directly addresses the local context: English, Urdu, and Roman Urdu mixed identification, with the interface automatically adjusting to right-to-left layout after language switching, and the output report fully generated in the target language. These three design points combined—action-oriented output, pragmatic model size trade-offs, and real language scenario coverage—give AI builders the most direct entry point for reference from this case.

Next time you’re designing an AI tool, we can start by asking a question: after the user gets the result, what’s their next action? Incorporating “follow-up recommendation steps” into the output design often does more to reduce the user’s actual risk than simply improving model accuracy.

📅 Source Information

🔗 Further Reading