Karya’s evaluation work includes large multilingual studies conducted with industry, academic, and public partners, as well as the development of open evaluation frameworks such as Samiksha and Pariksha. These efforts span tens of thousands of real-world queries, multiple Indian languages, and domains ranging from agriculture and law to healthcare.

Karya builds bespoke multi-modal datasets in Indic and other low-resource languages, enabling AI training across various formats:

We provide transcription across a range of audio and video sources, including interviews, conversational data, and domain-specific recordings. Work is carried out by verified contributors who are native speakers of the languages involved, ensuring transcripts reflect how people actually speak, code-switch, and use language in everyday contexts.

First-person, multimodal datasets built ethically at scale. Most datasets see the world from the outside. Egocentric data sees it from within. We work with communities to collect first-person, multimodal data — grounded in daily life, real environments, and local context. The result: AI systems trained on how the world is actually experienced.

Karya delivers high-quality data collection, annotation, and benchmarking services for NLP applications, ensuring accuracy and cultural relevance in:

Karya provides rich, localised datasets to train and evaluate AI-driven vision models, with: