---
title: "Lexa AI — Parth Kapoor"
description: "From-scratch transformer for English-to-Hindi/Telugu translation"
canonical: "https://parthkapoor.me/builds/lexa-ai"
---

# Lexa AI

> From-scratch transformer for English-to-Hindi/Telugu translation

Lexa AI is a multilingual machine translation model, built as an NLP course project. It implements a small encoder-decoder transformer from scratch in PyTorch, with no pretrained weights, to translate English into Hindi and Telugu.

A single encoder handles English input while the decoder is shared across both target languages, switched by a beginning-of-sentence token (<s-hi> or <s-te>). The model was trained on the Opus100 en-hi and en-te subsets using a custom BERT-style tokenizer, and is exposed through a Streamlit interface with a provided Dockerfile.

- Status: shipped
- Language: Python
- Started: Jun 2025
- Live: https://lexa.parthkapoor.me
- Source: https://github.com/parthkapoor-dev/lexa.ai

## Highlights

- 7M-parameter encoder-decoder transformer written from scratch, no pretrained weights
- Custom BERT-style tokenizer trained on the Opus100 en-hi/en-te dataset
- One encoder, one decoder shared across languages, switched via a BOS token per target language
- Streamlit UI for interactive testing, packaged with a Dockerfile for containerized runs

## Stack

- **Model**: PyTorch, Python
- **Data**: Opus100 dataset, Jupyter Notebook
- **Interface**: Streamlit, Docker

All builds: [https://parthkapoor.me/builds](https://parthkapoor.me/builds.md)