Anti-Scam Ninja

Anti Scam Ninja browser extention
Stop Elon Musk giveaway scams
Micheal Saylor impersonations
Scam Alert view: indicating a high score of content that is flagged as dangerous
Extension UI: scan youtube shorts and web pages service

Inspiration

I went after one of the biggest internet problems today which is financial scams. Every year people with untrained eye fall for financial scam abuse which leads to millions of lost funds. The perpetrators get better and better at creating schemes that lure people money into the hands of fraud. My chrome extension provides security for warning people of potential danger using chrome in built ai gemini nano model.

What it does

My project has youtube shorts integration which listens on changing video content. Once a video changed, it gets video id and pulls out its transcript. I pass the transcript into a gemini nano prompt api which is initialized on launching a browser extension and has initial prompt setup that provides a deep context of potential variations. It also follows a fix structure of response where I defined scam_score and explanation values which are later displayed on a new tab for a user. It only warns user if scam_score is higher than 0.9. Everything else is disregarded. If validation passes, users get appropriate warning.

How we built it

I set up a chrome extension with the help of react, vite and turbo monorepo library. I use tailwind for design and Google Prompt API gemini nano model which is locally download inside the browser. It also uses youtube-transcript library for pulling out transcript from youtube shorts which is later used with LLM in text analysis. Prompt API initialization gets initialPrompts with pre-trained scenario which helps to shape responses. The prompt also has a fixed scheme when it comes to provide response of scam analysis: scam_rating and explanation is used for displayed warning for users.

Here's the initial prompt: link

Challenges we ran into

Biggest challenge was to fine tune prompt initialization and provide useful content working with existing data. I ran into a lot of content which was crime based movies that made the model think users are getting persuaded with crime. I had to distinguish what is satire and what is potential fraudulent content that tries to lure people into financial scam.

Accomplishments that we're proud of

Fine tuning prompts of the model and playing with what youtube has to offer has been a fun job. It feels that internet has become a safer place once I have my extension turned on.

What we learned

Learned how thorough can be LLM responses with deep contextual analysis. The fraud vertical on the internet is an ever growing field and with emerging technologies like AI we need proper combat measures.

What's next for Anti-Scam Ninja

Next I'd like to focus on web page analysis, scanning for fraudulent text formations also certain indications that ask people send money to project wallets. More, Anti-Scam Ninja would go after twitter posts analysis impersonation.