Case study · Built and run by Foldox
sighting.ai
AI-content detection that explains each result.
- Kind
- AI-content detection
- Platforms
- Web app
Illustration with sample data.
The problem
People need to know whether something was made by AI, and why a tool thinks so. A bare score is hard to trust. And many checks send your content to someone else’s model.
What we built
- Detection for text, images, video and code, on self-hosted models.
- Reports that show their work: flagged passages, the signal that decided and a confidence range.
- Checks for content credentials (C2PA) and watermarks.
- Built to delete uploaded images and video once the analysis is done.
- Private reports, and a history people can clear at any time.
- A detection API, with keys people can revoke.
- Help answers from a hiy.ai support agent on the site.
Approach
Key decisions. What we chose, and why.
- 01
Self-hosted models
Detection runs on models we host, so content isn’t sent to a third-party detector.
- 02
A method for each kind of content
Text goes to a fine-tuned ModernBERT classifier, images to an ONNX image classifier, and code to a style analysis.
- 03
The reason, not just a score
Each report highlights the flagged passages, names the signal that decided and gives a confidence range.
- 04
Keep as little as possible
Uploaded images and video are deleted after analysis. The report and a fingerprint of the file are kept. The API can also leave text excerpts out of its reports.
Stack
What it runs on.
Web
- Next.js
- Vercel
Detection
- ModernBERT
- ONNX
- C2PA
Data and hosting
- Supabase
- Postgres
- Railway
Names are trademarks of their owners.
Capabilities shown
What this work shows. Each linked tag is a service we offer.
- Computer vision
- AI app development
- Text classification
- Self-hosted models
- Explainable results
Start a project
Need checks you can explain?
Tell us what you need to check, and why. We aim to reply within two working days.