ARCHIVES

ndia's digital economy is expanding rapidly, yet over ninety percent of automated customer support systems remain accessible only in English, excluding the majority of users who prefer native-language interaction. This paper presents the design, implementation, and empirical evaluation of an intelligent multilingual chatbot platform supporting six major Indian languages — Hindi, Bengali, Marathi, Tamil, Telugu, and Kannada — spanning both the Indo-Aryan and Dravidian language families. The system is built on a microservices architecture integrating FastText-based language identification, fine-tuned multilingual BERT (mBERT) and IndicBERT models for intent classification and named entity recognition, a Redis-backed context-aware multi-turn dialogue engine, AI4Bharat IndicConformer for automatic speech recognition (ASR), and Bhashini and Sarvam AI APIs for text-to-speech synthesis (TTS). An annotated e-commerce dataset of 18,550 examples across six languages and twenty intent categories was constructed. Three core experiments were conducted: (1) a comparative evaluation of mBERT versus IndicBERT for intent classification, where per-language IndicBERT models achieved a macro-averaged F1-score of 89.2%, outperforming mBERT (82.6%) by 6.6 percentage points; (2) a context-aware dialogue ablation study demonstrating a 7.8% improvement in multi-turn accuracy over the single-turn baseline; and (3) an ASR benchmark where IndicConformer achieved an average Word Error Rate of 13.4% across six languages, outperforming fine-tuned Whisper on Hindi (11.2% vs 13.8% WER). User acceptance testing with 25 participants yielded a CSAT score of 4.2/5.0 and an 84% task completion rate.