← all demos

on-device spellcheck

Type prose into the textarea below. Every time you stop typing for a moment, Dhamaka hands each word to an on-device masked language model running inside this browser tab and asks "what should go here?". Words the model considers unlikely in context are flagged. No rules, no hardcoded dictionary, no server — a real LLM reading your prose word by word.

Try a real sentence with typos:
Masked-LM spellcheck works best on real prose with real misspellings. Pure gibberish like asdsd qwdqd gets flagged correctly, but the suggestions for it will be nonsense too — there's no meaningful context for the model to predict from. That's a property of the algorithm, not a bug.

warming up the model…

First visit on this device downloads a ~65 MB masked language model (Xenova/distilbert-base-uncased). It's cached in your browser's IndexedDB forever after — every future visit is instant and works offline. 10–30 seconds on typical broadband, once.

—

The model runs through @huggingface/transformers, loaded lazily from esm.sh. Dhamaka wraps it behind the same task / SmartField / Transform API every other demo uses — the runtime underneath is pluggable, the product layer doesn't move.

draft

no issues yet

suggestions: 0 · source: — last call — ms

what's happening under the hood

  oninput (debounced 600ms) → SmartText → runTask("spellcheck", { eager: true })
       │
       ▼
  spellcheckTask.slow(text, context, engine)
       │
       ├─ tokenize input into words
       ├─ for each word:
       │     ├─ build "…prefix [MASK] suffix…"
       │     ├─ engine.fillMask(masked, top_k=20)  ← distilBERT via
       │     │                                        Transformers.js,
       │     │                                        runs in WASM
       │     └─ if original word not in top-20 → flag as misspelling,
       │        top predictions become corrections
       │
       └─ return structured { from, to, alternatives, index } list

  Nothing leaves the tab. No server, no API key, no rate limit.
  First visit downloads ~65 MB once, cached in IndexedDB forever.
  Per-call latency: ~100–300 ms per masked word on a laptop.

The formula demo still keeps its pattern rewrites (discounts, taxes, rounding, etc.) because those have objectively-correct structural answers and rules are a legitimate performance path there. Spellcheck is the opposite: probabilistic, context-dependent, long- tail. Rules there would contradict the thesis, so they're gone.

If your browser supports Chrome's window.ai Prompt API (Gemini Nano), Dhamaka will prefer that over Transformers.js — it's free, pre-downloaded, and GPU-accelerated. On every other browser you get Transformers.js. Same SDK, same task, same surface.