Demos
These are my own fine-tuned models, quantised to int8 and running in your browser through ONNX Runtime. Nothing you type is uploaded, because there is no server to upload it to.
Each model is about 67 MB and downloads only when you ask for it, then stays cached. The accuracy under each demo was measured on these exact quantised weights, not carried over from training.
Game identification
Paste a game review. The model names which of 1,232 games it is about, from the text alone.
60.0% top-1 across 1,232 titles, against a ~0.08% random baseline. Shipped as int8, down from 61.4% at fp32.
Review sentiment
Paste a game review. The model reads it as positive or negative.
87.3% on a balanced sample; 94.7% on the natural distribution, which is ~94% positive. Shipped as int8, identical to fp32.
How this works
Both models are DistilBERT, fine-tuned on 4.2 million game reviews. Exporting them to ONNX and quantising the weights to 8-bit integers cuts each one from about 268 MB to 67 MB, which is small enough to send to a browser.
Inference runs in a Web Worker so the download and the forward pass never freeze the page. The alternative was to host them on a server, which for a portfolio means a demo that sleeps after a day of no visitors and greets a recruiter with a cold start. Running on your device avoids that, and costs nothing to keep alive.
Quantisation is not free, so I measured it. Game identification lost 1.4 points going to int8, from 61.4% to 60.0%. Sentiment lost nothing at all.