Krishna Prasad Srinivasan

Krishna Prasad Srinivasan works on multilingual AI and document intelligence. He led the team behind Sarvam Vision, a three-billion-parameter vision-language model for English and 22 Indian languages. His work addresses a practical obstacle to useful AI: knowledge held in scanned records, historical collections, and regional-language documents must become readable by machines while preserving its meaning and structure.

From multilingual research to Sarvam Vision

Earlier in his career, Srinivasan was a researcher at Harvard, where he developed an OCR architecture using contrastive learning that outperformed industry benchmarks on complex multilingual documents. As Tech Lead for AI at Microsoft Research, he built multilingual education copilots and developed Indic translation models that outperformed commercial systems. In 2026, he led Sarvam’s vision models, research, and product as Head of Vision Models.

At Sarvam, he led a small team that trained Sarvam Vision, which launched in February 2026. Srinivasan reported that it led two global English OCR benchmarks at launch and the Indic OCR benchmark across 22 languages. Those results were time-bound: he acknowledged that subsequent releases had advanced the global leaderboard.

Srinivasan treats digitization as an upstream investment in AI. Much Indian-language knowledge exists in documents that have never been digitized; recovering it creates material that downstream models can learn from. The task requires more than recognizing characters. Layout and reading order must survive extraction so the document remains coherent and useful.

The choices behind his work

• Block-level OCR: Srinivasan and his team separated the model that reads individual document regions from the systems that determine layout and reading order. This divides recognizing a passage from determining its place in the document. Sarvam Vision’s document extraction capabilities include preserving layout and reading order and converting tables.

• State-space vision models: The team chose a state-space language backbone to reduce the cost of processing long visual sequences. It updates a running state token by token, avoiding the all-to-all attention computation that becomes expensive on dense pages. Srinivasan accepts some loss of recall as a tradeoff for lower inference costs, particularly with recognition organized around document blocks. His account of the architecture connects that choice to pages containing thousands of visual tokens.

• Language competence before visual training: The curriculum established a multilingual text foundation before image-text pretraining, supervised OCR fine-tuning, and reinforcement learning. A strong language prior helps resolve a smudged or ambiguous word. Srinivasan also insists on faithful transcription: when the original contains a mistake, OCR should preserve it rather than silently rewrite the document.

• Verifiable rewards and better training data: Character errors, table structure, equations, and grammar provide concrete signals for reinforcement learning. The team combines synthetic examples with real documents and uses evaluation results to improve training data. Srinivasan places particular weight on this data and evaluation machinery because many of the languages have little readily available labeled material. The training recipe alone does not solve that shortage. His explanation of training and evaluation makes these supporting systems central to the model’s performance.

Making document recognition usable

The product side extends into Akshar, Sarvam’s document workbench, which connects extracted text to locations in the original image and supports agent-assisted review. Confidence scores, block-level grounding, and proofreading give human reviewers a way to inspect uncertain passages against the source. Srinivasan reported deployments using Sarvam Vision to digitize more than 35 million pages for organizations including insurers, banks, governments, and historical preservation groups; this was work underway, rather than a claim that every page had been completed. Document workflows and deployments also inform his view of sovereignty: organizations need control over where document data goes and how it is used.

Sarvam’s model family subsequently expanded with Vision 2.1 in September 2026, adding structured extraction and Indic handwriting capabilities. Sarvam also published an OCR benchmark covering English and 22 Indian languages, with material dating from 1800 to the present. These company developments extend the scope of the document intelligence work beyond contemporary printed text to historical writing and more varied forms of extraction.