Scamless

Inspiration

Scams and fake content are everywhere right now — fake job offers, phishing links, AI-generated images used to deceive people. This one's personal: as a freelancer, I've received scam job offers myself, dressed up to look completely legitimate. Most people don't have a fast, reliable way to check if something's real before it's too late, and that gap is what Scamless is built to close.

What it does

Scamless is an AI-powered scam and fake-content detector covering the three most common ways scams reach people: text messages, links, and images. Paste a message, drop in a URL, or upload an image, and Scamless returns a clear risk score with an explanation — not just a black-box verdict.

How we built it

Backend (FastAPI): the text checker is a TF-IDF vectorizer + Logistic Regression model, trained on ~24,000 real examples combined from the Kaggle SMS Spam Collection and Phishing Email datasets, plus hand-written real scam examples. It hit 95% accuracy. The link checker does pure protocol-level checks — WHOIS lookups, SSL validation, redirect tracing — no ML, no external API. The image checker is a ResNet18 CNN, transfer-learned on the CIFAKE dataset (120k real vs. AI-generated images) using a free Colab GPU, reaching 96% test accuracy.

Frontend (React + Vite): a custom-designed landing page, and a /check page with a tabbed interface for all three live checkers, each calling the FastAPI backend in real time. Routed with React Router.

Challenges we ran into

Accomplishments that we're proud of

Getting all three detection models trained from scratch and working — not wrapped around someone else's API — is the thing I'm most proud of. The text classifier hit 95% accuracy on nearly 5,000 held-out real examples, and the image model hit 96% test accuracy after training on 120k images. Beyond the models, actually shipping a fully deployed, end-to-end working product (not just a local demo) by the deadline, after debugging through a lot of real infrastructure problems, felt like a genuine accomplishment.

What we learned

What's next for Scamless

Expand the video/deepfake detection feature that was scoped out for time Grow the text classifier's training data further, especially with more region-specific scam examples Add a browser extension so links and messages can be checked without copy-pasting into the site Improve the image detector to handle more image generation models as new ones emerge

Want me to adjust the tone anywhere, shorten any section, or add anything specific before you paste it in?

Built With

• computer-vision
• css
• fastapi
• google-colab
• javascript
• kaggle
• machine-learning
• natural-language-processing
• pandas
• python
• pytorch
• react
• react-router
• render
• resnet
• rest-api
• scikit-learn
• ssl
• tensorflow
• torchvision
• uvicorn
• vercel
• vite
• whois