CodeMixBridge is a Python-based machine learning project that improves English-to-Kannada language conversion by fine-tuning Google’s Gemma 4 E2B instruction model. It supports three tasks: English to Romanized Kannada, English to Kannada-script code-mixed text, and English to formal native Kannada. The project uses the AI4Bharat IndicCMix Kannada dataset, containing English sentences with corresponding Romanized, code-mixed, and native-script outputs. Data is normalized, deduplicated, grouped by source, and divided into deterministic training, validation, and testing sets without leakage. Each source record is expanded into three task-specific examples. The model is loaded in pre-quantized 4-bit format using Unsloth and bitsandbytes, allowing training on a Google Colab Tesla T4 GPU. LoRA parameter-efficient fine-tuning updates small adapter matrices while keeping the original Gemma weights frozen. PyTorch performs GPU computation, Hugging Face Transformers handles tokenization and generation, TRL provides the supervised fine-tuning trainer, and PEFT manages LoRA adapters. Training uses response-only loss, gradient accumulation, an 8-bit AdamW optimizer, and a cosine learning-rate schedule. Generic Gemma and fine-tuned Gemma are evaluated on identical held-out examples using deterministic generation and SacreBLEU chrF scores. Pandas organizes results, while Gradio provides an interactive interface for comparing both models’ Kannada outputs side by side.