A small neural model, running locally in your browser
Each of these runs in the box above
You pass the phrase. You get back the span, the value, the unit, the currency and a calibrated confidence
$ npm install gpu-sankhya
import { parse } from "gpu-sankhya";
const spans = parse("sava lakh ka phone");
// spans[0] → { span: "sava lakh", start: 0, end: 9,
// value: 125000, unit: "lakh", currency: null,
// confidence: 0.97, verified: true,
// tokens: [["PFX_SAVA","sava"],["UNIT_LAKH","lakh"]] }
// gate it: parse(text, { strict: true, minConfidence: 0.9 })
// bulk it: await parseBatch(texts, { backend: "webgpu" })
A char-level CNN tags every character; a deterministic arithmetic core does the maths, so the value can never be hallucinated
Characters to classes to a number, in 40,475 parameters