Ajamall family site · AI tools that run in your browser
← Back to Ajamall portal
Home › How it works

How it works

The tools on this site run AI models inside your own browser, not on a server. That is why your photos, audio and text are not uploaded.

1. Inference happens in the browser

  • The first time you run a tool, it downloads the model files (weights) and the programs it needs. Models mostly come from Hugging Face and programs from jsDelivr; some small model files come directly from this site.
  • After the download, the model computes on your input inside this browser. No server is involved in producing the result.
  • The programs that run the models are Transformers.js (Hugging Face) and ONNX Runtime Web. Text recognition uses Tesseract.js, PDFs use pdf.js, and face and pose detection use MediaPipe. All model and program licenses are on Sources & licenses.

2. WebGPU and wasm

If your browser supports WebGPU, the model can run on the graphics hardware, which may be faster. If it is not supported or not usable, the tool runs on wasm (WebAssembly, computed on the processor). The tool picks automatically, and wasm can take longer for the same model.

3. What you need

  • We recommend a recent desktop browser (Chrome, Edge, Firefox or Safari). Older browsers may not work.
  • Larger models need more memory and time. Large models such as text-generation models can be hundreds of MB up to more than 1 GB to download (the tool shows the size first), and on a device with little memory they can be slow or stall. If that happens, choose a smaller model.
  • To save memory, large images are scaled down to 2048 px on the long side.
  • You need an internet connection on the first run.

4. Size, storage and clearing it

Downloaded models are kept in your browser, so you do not have to download them again. To free the space, open your browser settings, find the clear site data (cookies and site data) menu, and remove the entry for this site (ai.ajamall.kr). The next run will download the model again. Menu names and locations differ a little between browsers.

5. Known limits

  • AI results can be wrong. Do not use them for important decisions (terms & disclaimer).
  • Text recognition (Tesseract) is fair on clean printed text but weak on handwriting and on skewed or low-contrast photos.
  • Translation (Opus-MT) supports specific language directions per tool. One tool cannot translate in every direction.
  • Speech to text (Whisper) understands many languages including Korean and Japanese, but smaller models are less accurate in Korean.
  • Small text-generation models (Qwen2.5) can be weak in Korean; larger ones are better but still limited on long text.
  • Noise reduction (RNNoise) suits steady noise and may dull the sound when music is present.
  • Image classification and object counting answer only within the English labels the model was trained on.

Last checked 2026-11-11