Research that ships.

Fine-tuning multilingual foundation models (NLLB) for translation between English and underserved Indian languages. Our evaluation framework uses held-out test sets co-built with native speakers, measuring BLEU, chrF, and human evaluation metrics.

Khasi machine translation is our flagship research track, currently shipping at BLEU 48.0 through Bha-Kha V4 on Bhasaflow, a Medharvix Systems platform.

Building structured evaluation pipelines for low-resource language AI. We treat evaluation as a first-class research problem: curating benchmarks, designing human evaluation protocols, and creating test sets that reflect real-world language use rather than synthetic data.

Developing text-to-speech and automatic speech recognition systems for languages with limited existing coverage. Our speech work addresses acoustic modelling, phonological adaptation, and prosody for languages like Khasi where standard speech models underperform.

Khasi TTS is in active research. Khasi ASR is under development.

Research on reading and understanding printed and handwritten text in Indian scripts. This work supports digitisation of institutional records, preservation of historical documents, and extraction of structured information from unstructured visual inputs.

Designing modular AI pipelines that combine translation, speech, and document understanding into integrated systems. This research underpins the Bhasaflow platform architecture and enables us to add new languages and modalities without rebuilding from scratch.