Krishna Prasad Srinivasan

Digitizing knowledge that models cannot yet read

Srinivasan describes digitization as an upstream problem for AI: knowledge preserved in documents is unavailable for training until machines can read it. His team’s three-billion-parameter model addresses complex Indian scripts and limited labeled data.[2]

Make extraction useful beyond plain text

Sarvam’s September 2026 Vision 2.1 release extends the document system with structured extraction and Indic handwriting recognition. It describes synthetic and real-world training data for forms, key-value pairs and handwriting. Layout parsing and reading-order prediction remain part of the surrounding system. The release also targets inconsistent outputs and hallucinations, alongside inference-stack changes intended to make production workloads cheaper. These are developments of the Sarvam team’s document program.[4]

Sarvam Vision uses a document harness around its vision-language model. Layout analysis and reading order organize the page, while the model handles recognition. Srinivasan explains the choice of a state-space backbone as a tradeoff: it carries a fixed-size state through a long sequence instead of comparing every token with every other token. That reduces sequence-processing costs while accepting some loss of recall. Training establishes multilingual text competence before adding visual inputs and OCR specialization.[2][3]