Privacy in AI detection has always felt like a choice between two bad options: upload your text to a stranger's server, or skip the check entirely. This project removes that trade-off by running both text and image detection entirely in the browser, with no upload required. That is a meaningful step forward, not because the models are new, but because the engineering around them respects the user's data by default.
The technical achievement here is compression without collapse. A 500 MB text model shrinks to 97 MB, and the image model lands at 26 MB, all while keeping inference fast enough for practical use, around 60 ms for a 100-token text check and 200 ms for an image on a laptop. The clever part is the mixed-precision quantization and the WiSE-FT interpolation that keeps the fine-tuned model from drifting too far from its original weights. This is the kind of pragmatic optimization that matters more than raw architecture novelty. It takes existing open detectors and makes them deployable in a context where privacy is the whole point. For readers building precision scoring models that skip text generation, this is a useful reference point: the same principle of reducing a model to its decision-making core applies here, but with the added constraint of running inside a browser sandbox.
The limitations are stated plainly, and they should be. A high score does not prove authorship. Short text, paraphrase, heavy editing, and newer generators all degrade accuracy. The first load is a heavy download, especially on phones. These are honest boundaries, and they matter because AI detection is already being over-trusted in classrooms and hiring pipelines. What this project does well is give users a tool that keeps their data local while making the model's uncertainty visible. That is a better foundation than a black-box API that returns a confident-looking percentage with no context.
What stands out is that all of this ran on a single laptop, training, tuning, quantization sweeps, and evaluation. That is a useful reminder that serious ML work does not require a cluster. It also opens a practical question: how far can this approach scale? The same techniques could apply to other detection tasks, like identifying machine-generated code or synthetic voices. The open question is whether quantization and client-side inference hold up as models grow more complex. For now, the takeaway is direct: if you care about privacy in AI tooling, this is a working proof that you do not have to sacrifice performance to get it. The next step is to stress-test it with your own text and see where it breaks. That is the kind of feedback this project is asking for, and it is the right way to push it forward.