All projects

Multilingual Tokenization Engine

A fairness-audited BPE tokenizer for English, Hindi, and Hinglish, with an explainability dashboard.

Tech stack

  • Python
  • FastAPI
  • Next.js
  • Streamlit
  • Scikit-learn
  • NLTK

The problem

Sub-word tokenizers often treat scripts unequally, fragmenting Hindi and Hinglish more aggressively than English. That imbalance quietly penalizes multilingual models — and it's rarely measured.

The solution

A BPE sub-word tokenizer with script-aware pre-tokenization for English, Hindi, and Hinglish, paired with a fairness-audit pipeline and an interactive explainability dashboard that surfaces tokenization parity across scripts.

Architecture

Script-aware pre-tokenization → BPE training → fairness-audit metrics → interactive dashboard (Streamlit/Next.js) over a FastAPI service. Robustness validated against typographic errors and OCR noise.

Highlights

  • Script-aware tokenization across English, Hindi, and Hinglish
  • Fairness-audit pipeline + interactive explainability dashboard
  • Stress-tested against typos and OCR noise for reliable inference