Hi HN, I’m Oscar, a Year 8 student from Australia who enjoys messing around with computers and AI.

I recently built an AI detector to build on my skills in computer science. I entered the prototype into the Oliphant Science Awards which is a local science competition (writing a 4000-word report on the methodology) and ended up winning, so I decided to polish it into a real web service that the world can make use of.

I noticed that schools and businesses are rushing to use AI detectors, but most commercial tools require you to send full, plaintext documents to cloud servers. For researchers or IP-sensitive work, sending data unencrypted to a third party (who might use it for training) is a major privacy risk. Additionally, current commercial AI detectors aren't very transparent and are unable to tell you why they come to a conclusion. I wanted to build something that helps people make informed decisions based on as much information as possible, not tell them a simple percentage with nothing to back it up.

I built Veredict to be secure and private: 1. The browser generates a one-time AES-256 key and encrypts the text locally using the Web Crypto API. 2. This AES key is encrypted using the server’s RSA public key. 3. The encrypted payload is sent to my backend (Python/FastAPI running on Modal serverless GPUs). 4. We decrypt in memory only for the split-second of inference. The plaintext is never saved to a database.

The detection logic uses an ensemble of 4 models (including statistical analysis of perplexity/burstiness and a fine-tuned BERT model) to output a confidence score.

A note on the login: I know HN prefers demos without sign-ups. Since I am a student paying for the GPU compute out of a limited budget, I strictly require Google Auth to prevent bots from draining my credits. I hope you understand.

The app provides a free daily quota (250 words) so you can test the architecture.

Link: https://veredictlabs.com

I'd really appreciate feedback on basically anything regarding my project :)

Show HN: Client-side encrypted AI detector using model ensembling