@pakhad/core

Gibberish and fake-input detection for text fields. Separates real names, words, and inputs from keyboard-mashed, bot-generated, and fake text.

Built for fintech signups, KYC, and anywhere you need to tell the grain from the chaff.

Install

npm install @pakhad/core @pakhad/locale-en

Quick Start

import { create } from '@pakhad/core';
import en from '@pakhad/locale-en';

const detector = create({ locales: [en] });

// Real name -> clean
detector.detect('Sarah Johnson', { fieldType: 'name' });
// { label: "clean", score: 0.082 }

// Keyboard mash -> gibberish
detector.detect('asdfgh qwerty', { fieldType: 'name' });
// { label: "gibberish", score: 0.785 }

How It Works

Pakhad runs 8 lightweight scorers against each token:

| Scorer | What it detects | |--------|----------------| | markov | Unlikely character sequences via n-gram model | | entropy | Abnormal character distribution | | keyboard_walk | QWERTY/AZERTY adjacency walks | | repetition | Run-length and pattern repetition | | vowel_consonant | Impossible vowel/consonant ratios | | name_list | Known name lookup via bloom filter | | numeric_pattern | Sequential/repeated digit patterns | | length_anomaly | Extremely short or long tokens |

Each scorer returns a score (0-1) and confidence (0-1). The ensemble combines them via confidence-weighted average per token, then length-weighted mean across tokens.

Multi-Language Support

import en from '@pakhad/locale-en';
import indiaLocales from '@pakhad/locale-in';

const detector = create({ locales: [en, ...indiaLocales] });

Field Type Detection

detector.detect('[email protected]');  // infers: email
detector.detect('cool_user123');       // infers: username
detector.detect('John Doe');           // infers: name
detector.detect('123 Main Street');    // infers: address

Custom Thresholds

detector.detect(input, {
  thresholds: { suspicious: 0.2, gibberish: 0.5 },
});

Custom Scorers

detector.registerScorer({
  name: 'my_scorer',
  defaultWeight: 0.3,
  score(token, ctx) {
    return { score: 0, confidence: 0 }; // silent by default
  },
});

Result Shape

interface DetectResult {
  score: number;              // 0 = clean, 1 = gibberish
  label: "clean" | "suspicious" | "gibberish";
  fieldType: { provided: FieldType | null; inferred: FieldType };
  locale: { mode: LocaleMode; script: string; candidates: string[]; matched: Record<string, string> };
  tokens: Array<{
    text: string;
    score: number;
    label: "clean" | "suspicious" | "gibberish";
    matchedLocale: string;
    scorers: Array<{ name: string; score: number; confidence: number; weight: number; details?: Record<string, unknown> }>;
  }>;
  warnings: string[];
  durationMs: number;
  version: string;
}

Benchmarks

| Metric | Value | |--------|-------| | Precision | 100% | | Recall | 89.8% | | F1 Score | 94.6% | | p99 Latency | 0.039ms |

Packages

| Package | Purpose | |---------|---------| | @pakhad/core | Detection engine (this package) | | @pakhad/locale-en | English locale (165k names, pre-built models) | | @pakhad/locale-in | Indian languages (Hindi, Marathi, Tamil, + 6 more) | | @pakhad/train | Build custom models from your own data |

Full Documentation

github.com/nikhilchintawar/pakhad

License

MIT