Pakistan Notice Helper

Pakistan Notice Helper is a bilingual safety assistant for confusing or suspicious Pakistani notices, bills, SMS messages, bank alerts, challans, and courier or customs messages. Paste text or upload a screenshot to receive:

• a risk label: Looks normal, Verify first, Suspicious, or Likely scam
• a simple English or Urdu explanation
• red flags, safe next steps, and a polite reply draft

The mobile-first interface has a persistent English/Urdu switch. Urdu mode uses a right-to-left layout and asks the live model to answer in Urdu. Roman Urdu and mixed-language inputs are also supported.

Pakistan Notice Helper checks common scam signals but does not provide official verification, legal advice, or financial advice. Verify through an official website or helpline before paying or sharing personal information.

Idea and Tech

This is a Backyard AI submission for the Build Small. Play Big. Hackathon. It applies a small model to a practical regional safety problem: helping people in Pakistan understand unfamiliar notices before they pay, reply, click a link, or share personal information.

The app combines a custom bilingual interface built on gradio.Server with Qwen3.5-4B Q8_0, served by a CUDA-enabled llama.cpp OpenAI-compatible endpoint on a Modal L4 GPU. It accepts text or screenshots, structures the model response into an actionable risk assessment, and can publish privacy-safe workflow traces only after explicit user consent.

Hackathon Fit

Demo and User Feedback

The demo video includes a real user using the app on a phone, taking screenshots of messages, and checking them with Pakistan Notice Helper.

The user especially liked the app's speed and the accuracy of its responses. They said they plan to use it in the future to check different scam or suspicious messages.

Architecture

Text or screenshot
        |
        v
Custom bilingual frontend on gradio.Server
        |
        v
Image preprocessing and structured safety prompt
        |
        v
Qwen3.5-4B Q8_0 MTP on llama.cpp
        |
        v
Modal L4 endpoint
        |
        v
Risk label, explanation, red flags, and safe next steps

Gradio provides queueing, API routes, and Hugging Face Spaces hosting without exposing a default Gradio UI. Live analyses always use the configured model endpoint. There is no heuristic fallback, and endpoint failures are shown explicitly. Bundled English examples are cached; Urdu analyses are generated live.

Model and Evaluation

• Model: unsloth/Qwen3.5-4B-MTP-GGUF using Qwen3.5-4B-Q8_0.gguf
• Runtime: llama.cpp OpenAI-compatible server
• Client: OpenAI Python SDK with a custom base_url
• Hosting: Modal L4 GPU
• Internal evaluation: 10/10 final regression cases, up from 9/10 initially

The focused test set covers banking notices, authority impersonation, prize scams, service messages, Urdu text, and screenshots. It is not a claim of general fraud-detection accuracy.

Run Locally

pip install -r requirements.txt
copy .env.example .env
python app.py

On macOS or Linux, use cp .env.example .env. Add MODAL_PROXY_KEY and MODAL_PROXY_SECRET to .env before starting live inference. Model URL, name, token limit, timeout, and retry settings are optional.

See docs/local_model_setup.md for full setup and deployment instructions.

Privacy-Safe Traces

Users can explicitly consent to sharing a minimized workflow trace. Traces may include language, input type, risk label, latency, response length, and a coarse error category.

They do not include raw notice text, OCR text, screenshots, names, phone numbers, credentials, or full model responses. See the trace dataset card.

Audit the live trace dataset with:

python -m traces.scripts.analyze_trace_dataset

Privacy and Limitations

• Inputs are processed in memory and are not written to disk by the app.
• Model requests are sent to the configured Modal endpoint.
• Redact CNIC numbers, account details, OTPs, PINs, and other sensitive data.
• Poor images, mixed scripts, abbreviations, or missing context can reduce reliability.
• Verify using contact details obtained independently from the message.

Official Verification

• PTA Complaint Management System
• PTA Numbering and Short Codes
• FIA Complaint Portal
• State Bank of Pakistan
• Federal Board of Revenue

Never rely on a verification link contained inside a suspicious message.