Inspiration

Scams in Pakistan (fake prize draws, fake JazzCash and Easypaisa messages, fake job offers) target parents and elders in the language they speak, and most detectors only work in English. Victims decide alone, under pressure. We wanted a tool that explains the scam in their language and lets them ask someone they trust.

What it does

ChaunkUp checks a pasted message or a screenshot and returns a verdict (Scam, Suspicious or Looks Safe) with a confidence score. It highlights the red flag phrases in the original text and explains why in English, Urdu and Roman Urdu. With one tap, "Alert my Circle" opens a WhatsApp or email link with the warning ready for a trusted family member. Nothing is ever sent automatically.

How we built it

We built it on DevMarketplace in six small modules: text analysis, languages and highlights, screenshot input, the Circle and alerts, history and polish, and hardening. We used Next.js, TypeScript and React, a PostgreSQL database, and the Gemini API for analysis. A SPEC.md, an AGENTS.md and a scam-analysis skill file guided the AI agent, so we could use few tokens.

Challenges we ran into

The hosting layer dropped our privacy cookie, so we switched to an anonymous per-browser ID sent with every request.
Gemini model names were retired while we built, which gave us 404 errors until we found the current model.
Making the AI return strict, validated structured output, with a safe fallback when it doesn't.

Accomplishments that we're proud of

Urdu script, Roman Urdu and English all work in one flow.
In a manual check, 10 of 10 test messages got the right verdict, from a 30-message test set included in the repo.
Contacts and history are private per browser, and the app handles bad input, unreadable images and AI failures without crashing.

What we learned

Small modules with clear plans, short prompts and tested steps keep AI-assisted development fast and cheap.

What's next for ChaunkUp

A WhatsApp bot so users can forward a message and get the verdict in chat, voice-note analysis, a shared scam-trends board, and more local languages (Sindhi, Pashto, Punjabi and Balochi).

Limitations

Data is private per browser with no accounts, so clearing browser data starts fresh. Gemini free-tier limits can slow repeated checks. Highlights depend on the AI returning matching phrases.

Built With

gemini-api
next.js
postgresql
react
typescript

Submitted to

Build The Next Big Thing with AI - Devmarketplace

Created by

I led ChaunkUp from idea to submission, and this was my first hackathon, full app from idea to submission, so I started without knowing how to plan, prompt or debug this way. I chose the scam-shield idea, wrote the SPEC.md and the plan files (PROJECT, ARCHITECTURE, MODULES, AGENTS and a scam-analysis skill), and split the build into six small modules so the AI agent on DevMarketplace could build one tested piece at a time and we could save tokens. I set up the database and the Gemini API key, wrote and sent the prompts, and tested every module myself with scam and genuine messages in English, Urdu and Roman Urdu. When things broke I read the logs and network requests to find the cause, such as a retired Gemini model that returned a 404, and the hosting layer dropping our privacy cookie, which I fixed by switching to a per-browser ID so users never see each other's contacts or history. I also ran the accuracy checks (10 of 10 verdicts correct on our test messages), managed our token use, and handled the GitHub repo, the README and the Devpost submission. Along the way I learned to write a spec, work in small tested steps, debug one problem at a time, and be honest about limitations.

Anoosha Hassan
Sana

Updates

Anoosha Hassan started this project — Oct 23, 2026 07:59 AM EDT

Leave feedback in the comments!

Log in or sign up for Devpost to join the conversation.