Project Sanmati is one of Karya’s longest-running and most impactful initiatives. Supported by the Gates Foundation, the project addresses a critical question for inclusive AI: how can women serve as both the workforce and the context-force in mitigating gender bias in Indian-language datasets?

Over the past two and a half years, we have worked with 20,000 low-income women across eight states to co-create frameworks for identifying gender bias across six Indian languages—Hindi, Marathi, Bhojpuri, Malayalam, Telugu, and Bengali. Through completing gender-focused tasks on the Karya platform, women acted as co-designers and evaluators in the data value chain, strengthening both data quality and community agency. This work resulted in a corpus of over 15,000 gender-aware sentences, which will underpin our next and final project phase: the development of a multilingual benchmarking tool for gender bias in Indian-language datasets.

Sanmati will be highlighted in a special panel, Women, Work, and the Future of AI, at the AI Impact Summit hosted by the Government of India in Delhi later this February.

As AI systems begin to shape everyday decisions across India, questions of usefulness, safety, and trust cannot be answered by abstract benchmarks alone. Samiksha was developed in response to this gap. It recognises that people most affected by model failures—those seeking guidance on health, education, finance, or legal matters—rarely have structured ways to participate in how AI systems are evaluated.

Samiksha is a multilingual evaluation pipeline that tests large language models against real, everyday Indian contexts, grounded in eleven Indian languages and four high-impact domains. It combines expert consultation, paid contributions from end users, and large-scale automated evaluation to surface where systems fall short in clarity, cultural fit, trustworthiness, and local relevance. By embedding lived experience directly into evaluation, Samiksha shifts assessment from laboratory assumptions to public life, contributing tools intended for use by researchers, policymakers, and model builders working in India’s AI ecosystem.

Samiksha will be showcased at the AI Impact Summit hosted by the Government of India in Delhi later this February.

Generative AI has expanded access to educational content, but it often struggles with cultural accuracy. In many contexts, particularly in the Global South, AI-generated stories reproduce settings, behaviours, or assumptions that feel unfamiliar or implausible to local learners. For educators, these gaps can undermine trust and reduce the effectiveness of teaching tools.

Kahani was developed by our long-time partners at Microsoft Research to address this challenge by keeping local practitioners at the centre of the creative process. The framework uses a staged, agent-based approach to storytelling, supported by human oversight at every step. Educators can intervene to refine language, gestures, settings, or visual details, ensuring stories remain grounded in local realities before they are finalised.

In 2025, a pilot in Kerala focused on financial literacy demonstrated the approach’s potential. Participants showed measurable improvements in learning outcomes, and all reported that the stories reflected their everyday experiences. The project lowered barriers for educators to create culturally resonant materials without specialised technical skills.

In 2025, the Indigenous Recipe Project culminated in the publication of a richly documented recipe book, From Hands that Feed, capturing culinary knowledge from communities across India in their own languages. Contributors recorded recipes alongside the practices, seasonal rhythms, and local reasoning that surround them—transforming everyday knowledge into a shared cultural record.

The project treated food as more than instruction. Each entry reflected how families preserve health, manage resources, and pass knowledge across generations. Contributors were paid for their work and credited by name, reinforcing the project’s commitment to ethical documentation.

The resulting book stands as both a cultural archive and a reminder that many forms of knowledge remain underrepresented in the datasets that shape technology today.