Krutrim

India’s biggest challenge in building strong multilingual LLMs has been the lack of clean, large-scale pretraining data in our languages. Our paper “BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages” has been accepted at AAAI-26. BhashaKritika introduces a 540B-token synthetic corpus across 10 Indic languages, built through a pipeline designed to generate high-quality, diverse text. It combines: • Generating content from real documents • Creating text using culturally relevant personas • Building math and reasoning material • Expanding coverage using topic-based retrieval • Translating high-quality English data into Indic languages A full quality-check pipeline filters for accuracy, fluency, language consistency, and bias. This corpus establishes one of the largest and most carefully filtered datasets for Indic languages, strengthening the foundation for future India-first LLMs. Read the full paper here: https://lnkd.in/d4grm7MR

Struggling with Indian accents in transcription? Hinglish breaking your speech-to-text tools? You’re not imagining it - most speech-to-text APIs simply weren’t built for how people in India actually talk. Shruti gives you transcripts that finally make sense for Indian English and Hinglish. With Shruti, you can: • Transcribe customer calls without replaying them again and again • Turn meetings and interviews into clean, easy-to-read notes • Create accurate transcripts, captions, and subtitles for videos and podcasts • Quickly process long recordings for audits, training, or documentation Upload your audio on Krutrim AI Studio and get a transcript that stays accurate even when the speaker switches between English and Hindi or speaks with a strong accent. Try Shruti on AI Studio: https://lnkd.in/dYDrPgnP More languages coming soon.

Voice data is everywhere — in customer calls, interviews, meetings, and videos. Yet most global ASR systems still falter when faced with Indian accents, dialects, or code-mixed speech. Accents that change every few kilometers, conversations that flow between Hindi and English, background sounds that global datasets rarely capture — these make speech recognition in India uniquely challenging. We built Shruti, Krutrim’s suite of Automatic Speech Recognition (ASR) models, to bridge that gap.  • Shruti-English-v1: for Indian English audio across diverse accents  • Shruti-Hinglish-MixedScript: outputs Hindi (Devanagari) + English (Roman)  • Shruti-Hinglish-Romanised: outputs both languages in Roman script Shruti helps enterprises transcribe customer interactions, meetings, and media content accurately — turning hours of audio into searchable, structured text. Available now on Krutrim AI Studio — ready to try in the playground or integrate via API.  👉 https://lnkd.in/dYDrPgnP

Multilingual models can be fluent, yet still miss the point culturally. The phrasing may be correct, but the tone, reasoning style, or shared context can feel off. This matters in real use — education, support, advice, everyday interaction. Today at EMNLP 2025, we’re presenting Pragyaan, our human-in-the-loop post-training data curation effort across 10 Indian languages. The goal is to build datasets that reflect how people actually speak, explain, instruct, and reason, not just how sentences translate. The pipeline blends: • Translation for high-quality coverage • Synthetic generation to broaden task and domain variety • Human refinement to ensure clarity, tone, and cultural relevance This supports more reliable instruction-following and alignment in multilingual contexts — where meaning depends on cultural grounding, not just linguistic accuracy. The attached poster outlines the pipeline, dataset composition, annotation workflow, and early downstream results. If you’re working on multilingual alignment, dataset design, or evaluation frameworks, we’re happy to exchange notes! Paper: https://lnkd.in/dRJr4EQB

Building Culturally Grounded Datasets for Indian Languages This Sunday at EMNLP 2025, we will present Pragyaan: Designing and Curating High-Quality Cultural Post-Training Datasets for Indian Languages. Multilingual models often miss nuance that is cultural, regional, and conversational. Pragyaan introduces a human-in-the-loop post-training pipeline to better capture how people across India speak, explain, instruct, and interact, across 10 languages. Key contributions:  • Pipeline using translation, synthetic generation, and human curation  • Scalable methods for capturing linguistic and cultural nuance  • Improved instruction alignment beyond literal correctness This work aims to strengthen multilingual alignment and evaluation practices for culturally diverse contexts. We look forward to discussing approaches at EMNLP and sharing further insights from our evaluations. Paper: https://lnkd.in/dRJr4EQB

Had an engaging discussion the other day at the panel on “Intelligence for India’s Next Leap”, alongside leaders from Zoho, Soket AI Labs, and Fujitsu Research. Thank you Fujitsu Research for the opportunity! We explored how India can build sovereign AI — covering both the opportunities and challenges in developing an end-to-end AI stack rooted in India’s context while remaining globally competitive. The conversation touched upon: 🇮🇳 Building sovereign AI: 🔸 India is one of the highest consumers of data, we also generate >20% of global data BUT bulk of this on apps and data centres that are outside India - this needs to change; we must build a full-stack sovereign AI ecosystem. 🔸 India uniquely offers rich and diverse data, cost-efficient innovation, and a strong talent base - strengths we need to tap into deliberately and strategically. 🔸 There are also challenges: availability of multilingual and Indian context data, evaluation benchmarks, data privacy, and our unique market dynamics - we discussed how organizations like Krutrim, Zoho, and Soket AI Labs are tackling these through foundational research and product innovation. 🏛️ Ecosystem collaboration: 🔸 Building sovereign AI is ambitious and resource intensive - We cannot do this in isolation. 🔸 The launch of technology focused missions like IndiaAI, shared compute capacity, data consolidation via AIKosh, and research funds by ANRF, RDI are promising efforts by the government. 🔸 There’s a growing need for enterprises and the VC ecosystem to adopt a patient capital mindset for deep-tech initiatives that have a longer gestation period. 🔸 Finally, startups, enterprises (local and global), and academia must forge meaningful, win-win collaborations aligned with our sovereign AI ambition. #SovereignAI #IndiaAI #Krutrim #AIforIndia #DeepTech #Innovation #AIecosystem

Take a look inside the Ola Gigafactory as our Founder, Bhavish Aggarwal explains the entire process of cell manufacturing. Watch the full interview here: https://lnkd.in/dBSkPfhn Chandra R. Srikanth moneycontrol.com

Take a look inside the Ola Gigafactory for the first time, as our Founder, Bhavish Aggarwal explains the entire process of cell manufacturing. Full interview here: https://lnkd.in/dEYjjaK5 Tamanna Inamdar NDTV Profit

Cloudera is now integrated into Krutrim’s AI-first datalake stack. This integration strengthens our vision of building a sovereign, AI native cloud for enterprises and empowers our advanced data lake capabilities with cost efficiency, security, and industry readiness