[
  {
    "startMs": 16640,
    "text": "All right. Um, hi everyone. Good"
  },
  {
    "startMs": 18480,
    "text": "afternoon. I'm Krishna. Uh, and I'm a"
  },
  {
    "startMs": 21279,
    "text": "general manager at Servin. Um today I'm"
  },
  {
    "startMs": 24800,
    "text": "going to tell you how a three billion"
  },
  {
    "startMs": 26960,
    "text": "parameter model uh small enough to run"
  },
  {
    "startMs": 30240,
    "text": "on a single GPU gets state-of-the-art uh"
  },
  {
    "startMs": 34079,
    "text": "in document AI and beats models 100"
  },
  {
    "startMs": 38000,
    "text": "times larger in size."
  },
  {
    "startMs": 41040,
    "text": "It is also a slightly unusual model in"
  },
  {
    "startMs": 44239,
    "text": "two ways. Firstly u the language model"
  },
  {
    "startMs": 46640,
    "text": "used is not a standard transformer. Um"
  },
  {
    "startMs": 49520,
    "text": "and the entire model building from data"
  },
  {
    "startMs": 52000,
    "text": "to training to compute uh was done end"
  },
  {
    "startMs": 55440,
    "text": "to end in India and for English and for"
  },
  {
    "startMs": 60079,
    "text": "22 official Indian languages uh which I"
  },
  {
    "startMs": 63199,
    "text": "think uh are personally uh uh one of the"
  },
  {
    "startMs": 65760,
    "text": "hardest document intelligence problems"
  },
  {
    "startMs": 67840,
    "text": "uh in the world right now. Um this is"
  },
  {
    "startMs": 70560,
    "text": "how we went from scratch to SOTA."
  },
  {
    "startMs": 74400,
    "text": "Who are we? Uh we are serv uh a"
  },
  {
    "startMs": 76720,
    "text": "sovereign foundation model company uh in"
  },
  {
    "startMs": 79840,
    "text": "India. We work across different"
  },
  {
    "startMs": 81920,
    "text": "modalities voice, text and vision. Uh in"
  },
  {
    "startMs": 85200,
    "text": "voice we have speech to text and"
  },
  {
    "startMs": 87759,
    "text": "texttospech models. In text we have uh a"
  },
  {
    "startMs": 91040,
    "text": "30 billion and 100 billion parameter uh"
  },
  {
    "startMs": 93759,
    "text": "uh models and in vision we have a"
  },
  {
    "startMs": 96400,
    "text": "document intelligence model which is"
  },
  {
    "startMs": 98159,
    "text": "what we are going to talk about today."
  },
  {
    "startMs": 101439,
    "text": "Um, India is largely missing from the"
  },
  {
    "startMs": 104880,
    "text": "machine readable world. Um, according to"
  },
  {
    "startMs": 107600,
    "text": "a published language distribution study,"
  },
  {
    "startMs": 110880,
    "text": "well under 1% of the common crawl uh"
  },
  {
    "startMs": 114000,
    "text": "corpus on which frontier models are"
  },
  {
    "startMs": 116079,
    "text": "trained uh have uh um Indian language"
  },
  {
    "startMs": 120079,
    "text": "representation."
  },
  {
    "startMs": 121600,
    "text": "Now in several forums you may have heard"
  },
  {
    "startMs": 123840,
    "text": "Frontier Labs say that um India is their"
  },
  {
    "startMs": 127119,
    "text": "um India is one of their larger uh uh"
  },
  {
    "startMs": 129599,
    "text": "and fastly growing markets and to that"
  },
  {
    "startMs": 133360,
    "text": "end um India is still missing uh in"
  },
  {
    "startMs": 136160,
    "text": "terms of data from frontier models that"
  },
  {
    "startMs": 139040,
    "text": "uh um are being trained every day here."
  },
  {
    "startMs": 142400,
    "text": "Now why is that? Uh primarily not"
  },
  {
    "startMs": 145040,
    "text": "because there is uh a lack of data or"
  },
  {
    "startMs": 148000,
    "text": "the knowledge does not exist. It is"
  },
  {
    "startMs": 150000,
    "text": "because the data has never been"
  },
  {
    "startMs": 151440,
    "text": "digitized and we are solving that"
  },
  {
    "startMs": 154560,
    "text": "problem at server."
  },
  {
    "startMs": 157680,
    "text": "Um why is Indic document intelligence"
  },
  {
    "startMs": 160879,
    "text": "hard? Uh firstly the goal is that of"
  },
  {
    "startMs": 164480,
    "text": "knowledge extraction not just purely"
  },
  {
    "startMs": 166640,
    "text": "text extraction. Uh extracting pure text"
  },
  {
    "startMs": 169599,
    "text": "without logical coherence is"
  },
  {
    "startMs": 171519,
    "text": "meaningless. And secondly in index"
  },
  {
    "startMs": 175360,
    "text": "scripts the shape of uh the words and"
  },
  {
    "startMs": 178800,
    "text": "the language you see are very different"
  },
  {
    "startMs": 181120,
    "text": "from what a machine sees. that is Indian"
  },
  {
    "startMs": 184159,
    "text": "languages have a complex uh set of uni"
  },
  {
    "startMs": 187120,
    "text": "codes fused together to get right the"
  },
  {
    "startMs": 189840,
    "text": "language model needs to be strong at all"
  },
  {
    "startMs": 192800,
    "text": "22 languages. Third, majority of Indian"
  },
  {
    "startMs": 196319,
    "text": "languages can be considered low resource"
  },
  {
    "startMs": 198400,
    "text": "due to which there is a lot of uh um"
  },
  {
    "startMs": 202080,
    "text": "lack in data today to train these"
  },
  {
    "startMs": 204080,
    "text": "models."
  },
  {
    "startMs": 206319,
    "text": "Our answer to this is server vision"
  },
  {
    "startMs": 208800,
    "text": "India's first sovereign vision language"
  },
  {
    "startMs": 210879,
    "text": "model built from scratch. um a three"
  },
  {
    "startMs": 214000,
    "text": "billion parameter model with state space"
  },
  {
    "startMs": 216400,
    "text": "architecture"
  },
  {
    "startMs": 218080,
    "text": "data compute and training all of which"
  },
  {
    "startMs": 220480,
    "text": "done in India. When we started the"
  },
  {
    "startMs": 223760,
    "text": "initial work in late 2025 most VLMs uh"
  },
  {
    "startMs": 227840,
    "text": "in the OCR space were monolith VLMs. uh"
  },
  {
    "startMs": 231200,
    "text": "they performed page level OCR and we at"
  },
  {
    "startMs": 235280,
    "text": "the time took a contrarian bet to focus"
  },
  {
    "startMs": 237599,
    "text": "on block level OCR instead adding a"
  },
  {
    "startMs": 240400,
    "text": "document hardness around the model. Um a"
  },
  {
    "startMs": 243760,
    "text": "lot of models since then uh released uh"
  },
  {
    "startMs": 246560,
    "text": "uh recently in 2026 have converged onto"
  },
  {
    "startMs": 249439,
    "text": "the same uh hardness plus small model"
  },
  {
    "startMs": 252080,
    "text": "paradigm for OCR which shows a lot of"
  },
  {
    "startMs": 254959,
    "text": "value in the direction we uh set off."
  },
  {
    "startMs": 258400,
    "text": "Serverom vision in particular uh has two"
  },
  {
    "startMs": 260799,
    "text": "hardness modules one for layout and one"
  },
  {
    "startMs": 262800,
    "text": "for reading hardness and a state space"
  },
  {
    "startMs": 266400,
    "text": "model uh VLM for block level OCR."
  },
  {
    "startMs": 271120,
    "text": "Again um today uh why do we use state"
  },
  {
    "startMs": 275520,
    "text": "space and not a transformer? Um most OCR"
  },
  {
    "startMs": 278320,
    "text": "models today uh like general VLMs, open"
  },
  {
    "startMs": 281440,
    "text": "source VLMs like Quen, Gemma etc. are"
  },
  {
    "startMs": 284560,
    "text": "all transformers based. We took an"
  },
  {
    "startMs": 286800,
    "text": "alternate approach uh using an SSM. Why?"
  },
  {
    "startMs": 291280,
    "text": "Both are fundamentally sequenced models"
  },
  {
    "startMs": 293280,
    "text": "transformers and SSMs but with very"
  },
  {
    "startMs": 296000,
    "text": "different underlying mechanics. A"
  },
  {
    "startMs": 298800,
    "text": "transformer has every token look at"
  },
  {
    "startMs": 301280,
    "text": "every other token where the compute"
  },
  {
    "startMs": 303759,
    "text": "grows with square of the sequence"
  },
  {
    "startMs": 306800,
    "text": "length. Uh that is the L cross L"
  },
  {
    "startMs": 309919,
    "text": "interactions and the memory goes as the"
  },
  {
    "startMs": 312800,
    "text": "sequence grows as well. Now on the other"
  },
  {
    "startMs": 315440,
    "text": "hand, SSMs have a single state. They"
  },
  {
    "startMs": 318479,
    "text": "maintain a single state through the"
  },
  {
    "startMs": 320560,
    "text": "sequence updating it token by token. Now"
  },
  {
    "startMs": 323680,
    "text": "as compute grows only linearly, the"
  },
  {
    "startMs": 326080,
    "text": "memory stays constant for SSMS. Now why"
  },
  {
    "startMs": 330320,
    "text": "is this the right uh architecture for"
  },
  {
    "startMs": 332880,
    "text": "OCR? It really depends on the uh uh"
  },
  {
    "startMs": 336080,
    "text": "trade-offs here uh for long documents in"
  },
  {
    "startMs": 338560,
    "text": "particular which can lead up to uh uh 5"
  },
  {
    "startMs": 341759,
    "text": "to 10,000 uh visual tokens per page and"
  },
  {
    "startMs": 346160,
    "text": "quadratic complexity of you know uh"
  },
  {
    "startMs": 348880,
    "text": "compute and memory becomes very"
  },
  {
    "startMs": 350720,
    "text": "expensive for inference. On the other"
  },
  {
    "startMs": 352960,
    "text": "hand,"
  },
  {
    "startMs": 354639,
    "text": "um doing block level OCR with somewhat"
  },
  {
    "startMs": 357680,
    "text": "of a lossy recall using SSMs can be"
  },
  {
    "startMs": 360320,
    "text": "justifi justified to uh um avoid high"
  },
  {
    "startMs": 364800,
    "text": "compute costs imposed by transformers."
  },
  {
    "startMs": 368960,
    "text": "So how do you actually train this? We"
  },
  {
    "startMs": 372000,
    "text": "built a staged curriculum four stages"
  },
  {
    "startMs": 374720,
    "text": "each uh building on the last um stage"
  },
  {
    "startMs": 378240,
    "text": "one is texton pre-training uh 13"
  },
  {
    "startMs": 381280,
    "text": "trillion tokens across English, Indian"
  },
  {
    "startMs": 383919,
    "text": "language, text, math and code. This"
  },
  {
    "startMs": 386960,
    "text": "builds the three billion parameters"
  },
  {
    "startMs": 388960,
    "text": "language backbone and a strong language"
  },
  {
    "startMs": 392720,
    "text": "prior is what lets the model resolve a"
  },
  {
    "startMs": 395520,
    "text": "smudged or an ambiguous text from the"
  },
  {
    "startMs": 398639,
    "text": "image. the same way you can read a half"
  },
  {
    "startMs": 401520,
    "text": "blurred word because you know which word"
  },
  {
    "startMs": 403360,
    "text": "is supposed to uh u be at the right"
  },
  {
    "startMs": 406400,
    "text": "place. So we focused on building the"
  },
  {
    "startMs": 409280,
    "text": "competence of the language model first"
  },
  {
    "startMs": 412000,
    "text": "uh before the model ever even sees a"
  },
  {
    "startMs": 414319,
    "text": "pixel. Stage two is where we performed"
  },
  {
    "startMs": 418479,
    "text": "continual pre-training uh on 300 million"
  },
  {
    "startMs": 421280,
    "text": "image text pairs. This teaches the"
  },
  {
    "startMs": 423440,
    "text": "language model general vision"
  },
  {
    "startMs": 425680,
    "text": "capabilities and teaches how to see how"
  },
  {
    "startMs": 427599,
    "text": "to interpret pixels and so on. Then"
  },
  {
    "startMs": 430319,
    "text": "comes stage three where we performed uh"
  },
  {
    "startMs": 432880,
    "text": "supervised fine-tuning on 100 uh uh"
  },
  {
    "startMs": 435440,
    "text": "million uh OCR uh samples. Now the um"
  },
  {
    "startMs": 441840,
    "text": "stage three is primarily focused on"
  },
  {
    "startMs": 444000,
    "text": "letting a general VM model become strong"
  },
  {
    "startMs": 446639,
    "text": "at OCR. What that includes is having"
  },
  {
    "startMs": 450319,
    "text": "diverse data across all 22 languages and"
  },
  {
    "startMs": 453120,
    "text": "English and also incorporating all kinds"
  },
  {
    "startMs": 455759,
    "text": "of uh uh document components such as"
  },
  {
    "startMs": 459039,
    "text": "tables and uh um equations and"
  },
  {
    "startMs": 462800,
    "text": "handwritten documents and so forth. Now"
  },
  {
    "startMs": 465759,
    "text": "and stage four is re reinforcement"
  },
  {
    "startMs": 467840,
    "text": "learning which helps push past the"
  },
  {
    "startMs": 470639,
    "text": "ceiling of whatever uh supervised"
  },
  {
    "startMs": 473120,
    "text": "fine-tuning is able to achieve. Here you"
  },
  {
    "startMs": 475919,
    "text": "will see this is a standard recipe."
  },
  {
    "startMs": 478560,
    "text": "However, the moat here is the two things"
  },
  {
    "startMs": 482639,
    "text": "underneath the data layer and the eval"
  },
  {
    "startMs": 485599,
    "text": "layer. Um the first is the data engine."
  },
  {
    "startMs": 489039,
    "text": "For most of the 22 languages, there is"
  },
  {
    "startMs": 491039,
    "text": "no label data available readily."
  },
  {
    "startMs": 493919,
    "text": "When label data does not exist, building"
  },
  {
    "startMs": 497039,
    "text": "the data engine becomes hard. And that"
  },
  {
    "startMs": 500000,
    "text": "is something we have done extensively."
  },
  {
    "startMs": 502240,
    "text": "We've built pipelines to create"
  },
  {
    "startMs": 504240,
    "text": "synthetic data and doc data from real"
  },
  {
    "startMs": 506800,
    "text": "documents and also helped"
  },
  {
    "startMs": 510160,
    "text": "build the uh pipeline for continuously"
  },
  {
    "startMs": 513599,
    "text": "improving the data that goes into"
  },
  {
    "startMs": 516000,
    "text": "training based on the eval performance"
  },
  {
    "startMs": 518000,
    "text": "and so on. So we are currently actively"
  },
  {
    "startMs": 520640,
    "text": "looking at the paradigm of RLMs um as uh"
  },
  {
    "startMs": 524880,
    "text": "uh we explore the agentic vision"
  },
  {
    "startMs": 526720,
    "text": "capabilities for our models in the"
  },
  {
    "startMs": 528720,
    "text": "upcoming releases. The second mode is"
  },
  {
    "startMs": 531279,
    "text": "the eval. You cannot reach"
  },
  {
    "startMs": 533519,
    "text": "state-of-the-art if uh um you cannot"
  },
  {
    "startMs": 536720,
    "text": "measure how well your model performs. We"
  },
  {
    "startMs": 539440,
    "text": "have curated extensive evals to ensure"
  },
  {
    "startMs": 542240,
    "text": "what we are measuring is truly sota and"
  },
  {
    "startMs": 544560,
    "text": "also truly meaningful in terms of"
  },
  {
    "startMs": 546640,
    "text": "usefulness for end uh uh users. So I"
  },
  {
    "startMs": 551519,
    "text": "want to take just a minute to dwell on"
  },
  {
    "startMs": 553680,
    "text": "our RL pipeline uh stage four because uh"
  },
  {
    "startMs": 558080,
    "text": "a lot of gain comes from RL. Essentially"
  },
  {
    "startMs": 561040,
    "text": "in OCR correctness is a machine readable"
  },
  {
    "startMs": 564240,
    "text": "problem right a lot of tests can be set"
  },
  {
    "startMs": 566959,
    "text": "up to uh um reward and score the models"
  },
  {
    "startMs": 571200,
    "text": "against uh uh these uh samples that you"
  },
  {
    "startMs": 574480,
    "text": "create. And in the world of"
  },
  {
    "startMs": 577040,
    "text": "deterministic OCR, these are all machine"
  },
  {
    "startMs": 579839,
    "text": "uh uh verifiable. Hence, RL gives us a"
  },
  {
    "startMs": 583519,
    "text": "major boost."
  },
  {
    "startMs": 585440,
    "text": "Sample a group, score with unit tests,"
  },
  {
    "startMs": 587920,
    "text": "reinforce what the average uh uh value"
  },
  {
    "startMs": 591519,
    "text": "to beat is, and then repeat the process."
  },
  {
    "startMs": 594160,
    "text": "And this makes RLVR for OCR very very"
  },
  {
    "startMs": 596959,
    "text": "scalable."
  },
  {
    "startMs": 598880,
    "text": "Now after all the from scratch training"
  },
  {
    "startMs": 601519,
    "text": "we've done and the data effort we've"
  },
  {
    "startMs": 603519,
    "text": "made we were able to establish sot in"
  },
  {
    "startMs": 606240,
    "text": "two of uh the global English benchmarks"
  },
  {
    "startMs": 609760,
    "text": "one is CRB bench and the other is"
  },
  {
    "startMs": 612240,
    "text": "omnidoc bench at launch we were at 84.3"
  },
  {
    "startMs": 616000,
    "text": "on CR and then 93.2 into on omnidok"
  },
  {
    "startMs": 619040,
    "text": "bench. Um the models uh that have"
  },
  {
    "startMs": 622720,
    "text": "released since have moved the needle"
  },
  {
    "startMs": 624720,
    "text": "quite a bit and we'll soon have a a"
  },
  {
    "startMs": 626959,
    "text": "stronger model in the global leaderboard"
  },
  {
    "startMs": 629120,
    "text": "as well."
  },
  {
    "startMs": 631279,
    "text": "Secondly and more importantly um on 22"
  },
  {
    "startMs": 634800,
    "text": "Indian languages we have an unbeaten"
  },
  {
    "startMs": 637279,
    "text": "lead even compared to all the frontier"
  },
  {
    "startMs": 639279,
    "text": "models such as Gemini and uh chat GPT"
  },
  {
    "startMs": 642959,
    "text": "and Opus and so on. And this is where"
  },
  {
    "startMs": 645360,
    "text": "we've extended the lead significantly"
  },
  {
    "startMs": 647279,
    "text": "and we remain uh uh strong compared to"
  },
  {
    "startMs": 650079,
    "text": "all of these newer models that have come"
  },
  {
    "startMs": 652160,
    "text": "in."
  },
  {
    "startMs": 653680,
    "text": "Now server vision pars our agentic"
  },
  {
    "startMs": 657040,
    "text": "document intelligence workbench called"
  },
  {
    "startMs": 659279,
    "text": "aure where we enable"
  },
  {
    "startMs": 663120,
    "text": "human in the loop agentic digitizations"
  },
  {
    "startMs": 666079,
    "text": "uh and uh extractions and then"
  },
  {
    "startMs": 668640,
    "text": "injections for various uh downstream uh"
  },
  {
    "startMs": 671120,
    "text": "document intelligence problems. We"
  },
  {
    "startMs": 673200,
    "text": "provide confidence scores um and we uh"
  },
  {
    "startMs": 676399,
    "text": "have block level grounding and then also"
  },
  {
    "startMs": 678399,
    "text": "the ability to do uh agentic proof"
  },
  {
    "startMs": 680800,
    "text": "reading etc."
  },
  {
    "startMs": 683360,
    "text": "Now benchmarks and soda are uh uh one"
  },
  {
    "startMs": 687279,
    "text": "thing they have their place. Today some"
  },
  {
    "startMs": 689600,
    "text": "of the largest enterprises in the world"
  },
  {
    "startMs": 691440,
    "text": "from insurance to banking to governments"
  },
  {
    "startMs": 694560,
    "text": "to historical preservation organizations"
  },
  {
    "startMs": 697760,
    "text": "are using server vision to digitize more"
  },
  {
    "startMs": 700399,
    "text": "than 35 million pages across English and"
  },
  {
    "startMs": 703600,
    "text": "uh uh 22 Indian languages. The model is"
  },
  {
    "startMs": 707360,
    "text": "available as API and on prem and then"
  },
  {
    "startMs": 711200,
    "text": "the agentic platform."
  },
  {
    "startMs": 714480,
    "text": "So in conclusion until 4 months ago uh"
  },
  {
    "startMs": 718560,
    "text": "there was no sovereign model uh from"
  },
  {
    "startMs": 720640,
    "text": "India. Today we have server vision which"
  },
  {
    "startMs": 723440,
    "text": "was trained from scratch and has"
  },
  {
    "startMs": 725360,
    "text": "established sot um at a price point that"
  },
  {
    "startMs": 728959,
    "text": "is extremely competitive with uh all the"
  },
  {
    "startMs": 731839,
    "text": "other solutions including open and close"
  },
  {
    "startMs": 734160,
    "text": "source. Uh first we've started by uh uh"
  },
  {
    "startMs": 737600,
    "text": "solving some of the hardest problems for"
  },
  {
    "startMs": 739920,
    "text": "Indian language document intelligence."
  },
  {
    "startMs": 742320,
    "text": "Soon uh we'll be releasing general"
  },
  {
    "startMs": 744240,
    "text": "purpose uh VLMs which are capable of"
  },
  {
    "startMs": 746959,
    "text": "much more uh uh vision capabilities and"
  },
  {
    "startMs": 750480,
    "text": "uh look forward to uh uh all of you"
  },
  {
    "startMs": 752800,
    "text": "trying out our models. Thank you."
  },
  {
    "startMs": 756533,
    "text": "[applause]"
  },
  {
    "startMs": 759600,
    "text": "Happy to answer any questions. Yeah."
  },
  {
    "startMs": 768320,
    "text": "Yes."
  },
  {
    "startMs": 769600,
    "text": ">> Yes. Absolutely. Uh in general, the"
  },
  {
    "startMs": 771600,
    "text": "language capability of the 22 languages"
  },
  {
    "startMs": 774240,
    "text": "enhances English uh by quite a margin."
  },
  {
    "startMs": 777680,
    "text": "And this is applicable to any low"
  },
  {
    "startMs": 779839,
    "text": "resource language, not just Indian low"
  },
  {
    "startMs": 782240,
    "text": "resource languages."
  },
  {
    "startMs": 785600,
    "text": ">> Yeah."
  },
  {
    "startMs": 807760,
    "text": "Right. Um we don't quite go in that"
  },
  {
    "startMs": 810079,
    "text": "direction with this model because this"
  },
  {
    "startMs": 811920,
    "text": "is a vision focused uh uh model wherein"
  },
  {
    "startMs": 814959,
    "text": "we are focused on extracting information"
  },
  {
    "startMs": 816959,
    "text": "or knowledge from documents. But yes,"
  },
  {
    "startMs": 819440,
    "text": "there could be uh some parallels to um"
  },
  {
    "startMs": 822240,
    "text": "helping models use general languages to"
  },
  {
    "startMs": 826000,
    "text": "speed up coding as well. But yeah,"
  },
  {
    "startMs": 827839,
    "text": "that's not the uh periphery of this uh"
  },
  {
    "startMs": 830959,
    "text": "uh work."
  },
  {
    "startMs": 833440,
    "text": "Yeah."
  },
  {
    "startMs": 850160,
    "text": "Right. Um so there are two things uh we"
  },
  {
    "startMs": 853440,
    "text": "uh create artificial documents synthetic"
  },
  {
    "startMs": 855920,
    "text": "documents as they are called for general"
  },
  {
    "startMs": 858399,
    "text": "post-raining uh that includes SFT and"
  },
  {
    "startMs": 861360,
    "text": "RL. Um however coming to your specific"
  },
  {
    "startMs": 864560,
    "text": "question on RL uh we don't uh you can"
  },
  {
    "startMs": 867519,
    "text": "generate uh synthetic documents there as"
  },
  {
    "startMs": 869440,
    "text": "well. However, um the best thing to do"
  },
  {
    "startMs": 872000,
    "text": "is to take real world documents complex"
  },
  {
    "startMs": 874079,
    "text": "enough that uh help you set up unit"
  },
  {
    "startMs": 876959,
    "text": "tests or different kinds of uh uh"
  },
  {
    "startMs": 879920,
    "text": "rewards, right? like say a a reward"
  },
  {
    "startMs": 882560,
    "text": "based on character error rate or a a"
  },
  {
    "startMs": 885040,
    "text": "reward based on table structure or a"
  },
  {
    "startMs": 887199,
    "text": "math equation or something uh that is"
  },
  {
    "startMs": 889839,
    "text": "pertinent to a language or like grammar"
  },
  {
    "startMs": 892399,
    "text": "rewards or so on and then help the model"
  },
  {
    "startMs": 895120,
    "text": "iteratively improve based on uh um rules"
  },
  {
    "startMs": 898079,
    "text": "that the model is able to produce at"
  },
  {
    "startMs": 899600,
    "text": "different uh u settings. Yeah. Yes."
  },
  {
    "startMs": 921440,
    "text": "Sorry, did you say big fan of Chundra?"
  },
  {
    "startMs": 924240,
    "text": ">> All right. Uh, that's from a different"
  },
  {
    "startMs": 925760,
    "text": "lab. Uh, I'm also a big fan of the lab"
  },
  {
    "startMs": 928720,
    "text": "that produced Chundra. But yes, took"
  },
  {
    "startMs": 930560,
    "text": "your question uh on Indic benchmark."
  },
  {
    "startMs": 932880,
    "text": "Yes, we will release the server indic"
  },
  {
    "startMs": 934959,
    "text": "benchmark that we have created for 22"
  },
  {
    "startMs": 938079,
    "text": "languages and it uh spans um a a huge"
  },
  {
    "startMs": 942480,
    "text": "time period starting from 1800s until"
  },
  {
    "startMs": 945040,
    "text": "modern day and also different kinds of"
  },
  {
    "startMs": 946959,
    "text": "layouts, different kinds of documents in"
  },
  {
    "startMs": 949199,
    "text": "Indian languages. You can imagine"
  },
  {
    "startMs": 950800,
    "text": "documents with pros, poetry, literature,"
  },
  {
    "startMs": 953920,
    "text": "uh tables, finance and all of those"
  },
  {
    "startMs": 955759,
    "text": "things. We'll soon release that"
  },
  {
    "startMs": 957279,
    "text": "benchmark uh in public as well."
  },
  {
    "startMs": 985440,
    "text": "So translation is again uh"
  },
  {
    "startMs": 987600,
    "text": "transliteration is again not directly"
  },
  {
    "startMs": 990000,
    "text": "involved here because in OCR what we"
  },
  {
    "startMs": 992560,
    "text": "care about is high fidelity extraction"
  },
  {
    "startMs": 995279,
    "text": "right uh you want even if there is a"
  },
  {
    "startMs": 997199,
    "text": "mistake in an image you want that"
  },
  {
    "startMs": 998800,
    "text": "mistake to be extracted correctly and"
  },
  {
    "startMs": 1000480,
    "text": "not the model uh uh making changes at"
  },
  {
    "startMs": 1003360,
    "text": "whim uh hence that is not directly"
  },
  {
    "startMs": 1005839,
    "text": "applicable but yes uh what we see is a"
  },
  {
    "startMs": 1008240,
    "text": "lot of data that is coming through"
  },
  {
    "startMs": 1010000,
    "text": "transliteration for even OCR training"
  },
  {
    "startMs": 1012160,
    "text": "now and the um jury is out on what is"
  },
  {
    "startMs": 1016240,
    "text": "the quality of that data and how useful"
  },
  {
    "startMs": 1018560,
    "text": "it is etc"
  },
  {
    "startMs": 1020800,
    "text": "Yeah."
  },
  {
    "startMs": 1047760,
    "text": "Right. Um there are few things to"
  },
  {
    "startMs": 1049919,
    "text": "consider. Um first is no other model"
  },
  {
    "startMs": 1053760,
    "text": "frontier uh closed source or open model"
  },
  {
    "startMs": 1056559,
    "text": "does well with complex Indian language"
  },
  {
    "startMs": 1059360,
    "text": "documents right so for a size of uh for"
  },
  {
    "startMs": 1063840,
    "text": "a country with 1.4 4 billion people you"
  },
  {
    "startMs": 1067280,
    "text": "want to be able to address their"
  },
  {
    "startMs": 1069200,
    "text": "day-to-day lives in uh uh terms of"
  },
  {
    "startMs": 1071679,
    "text": "document intelligence problems because"
  },
  {
    "startMs": 1073840,
    "text": "there is a lot of paperwork in India"
  },
  {
    "startMs": 1075919,
    "text": "right it is a up and cominging country"
  },
  {
    "startMs": 1078400,
    "text": "and it's still being digitized as we"
  },
  {
    "startMs": 1080400,
    "text": "speak and it's important that"
  },
  {
    "startMs": 1082559,
    "text": "representation or ability to digitize"
  },
  {
    "startMs": 1084559,
    "text": "the country exists to begin with."
  },
  {
    "startMs": 1086960,
    "text": "Secondly on uh the training side what we"
  },
  {
    "startMs": 1090240,
    "text": "have uh uh done in particular is to"
  },
  {
    "startMs": 1094640,
    "text": "create data that can be added to a lot"
  },
  {
    "startMs": 1098000,
    "text": "of training downstream. So uh we are at"
  },
  {
    "startMs": 1100799,
    "text": "a very early stage in terms of AI"
  },
  {
    "startMs": 1102880,
    "text": "becoming you know a regular parts of our"
  },
  {
    "startMs": 1106160,
    "text": "lives and we need the data to start to"
  },
  {
    "startMs": 1109039,
    "text": "be able to eventually get to a place"
  },
  {
    "startMs": 1110960,
    "text": "where we can uh have personalized agents"
  },
  {
    "startMs": 1114080,
    "text": "in the language you prefer in language I"
  },
  {
    "startMs": 1116080,
    "text": "prefer and however I want in uh uh in"
  },
  {
    "startMs": 1119039,
    "text": "some so to do all of that we need to"
  },
  {
    "startMs": 1121440,
    "text": "start somewhere and the data needs to be"
  },
  {
    "startMs": 1123440,
    "text": "created and if the data is of good"
  },
  {
    "startMs": 1125280,
    "text": "quality then that helps the model become"
  },
  {
    "startMs": 1127840,
    "text": "sota and If a model is sot then uh from"
  },
  {
    "startMs": 1132000,
    "text": "uh insurance agencies to government"
  },
  {
    "startMs": 1134240,
    "text": "organizations to others who care about"
  },
  {
    "startMs": 1137039,
    "text": "sovereignity in AI right uh as a"
  },
  {
    "startMs": 1139679,
    "text": "government body I cannot have a model"
  },
  {
    "startMs": 1142000,
    "text": "deployed elsewhere and I don't know"
  },
  {
    "startMs": 1143679,
    "text": "where the data is going for uh uh the"
  },
  {
    "startMs": 1146720,
    "text": "transcriptions so I need to be able to"
  },
  {
    "startMs": 1149039,
    "text": "control where the data is sent and when"
  },
  {
    "startMs": 1151600,
    "text": "it is used how much it is used etc. So"
  },
  {
    "startMs": 1154480,
    "text": "sovereignity becomes very crucial and"
  },
  {
    "startMs": 1156320,
    "text": "hence this model really is now uh uh"
  },
  {
    "startMs": 1158960,
    "text": "like it's been what four months since we"
  },
  {
    "startMs": 1160720,
    "text": "launched and already we are we are"
  },
  {
    "startMs": 1162960,
    "text": "digitizing 35 million pages. So that is"
  },
  {
    "startMs": 1165520,
    "text": "that shows that the market was waiting"
  },
  {
    "startMs": 1167120,
    "text": "for something sovereign in this space"
  },
  {
    "startMs": 1168640,
    "text": "that can really kickstart uh the AI"
  },
  {
    "startMs": 1171840,
    "text": "digitization wave in India. So yeah"
  },
  {
    "startMs": 1174080,
    "text": "primarily like three aspects one on uh"
  },
  {
    "startMs": 1176400,
    "text": "sovereignity the other is the model uh"
  },
  {
    "startMs": 1179280,
    "text": "capability itself and the third is data"
  },
  {
    "startMs": 1181679,
    "text": "uh that is required to train these"
  },
  {
    "startMs": 1183200,
    "text": "models."
  },
  {
    "startMs": 1194320,
    "text": "Yeah, it is it is uh um 40% uh indic and"
  },
  {
    "startMs": 1199120,
    "text": "then uh the rest is English and"
  },
  {
    "startMs": 1201039,
    "text": "comprises of math and code etc."
  },
  {
    "startMs": 1208080,
    "text": ">> All right."
  },
  {
    "startMs": 1209679,
    "text": "Just a curious question."
  },
  {
    "startMs": 1224960,
    "text": "Um currently the uh the deployments"
  },
  {
    "startMs": 1228159,
    "text": "we've made uh with this model are in"
  },
  {
    "startMs": 1230559,
    "text": "different states uh uh trying to"
  },
  {
    "startMs": 1233120,
    "text": "digitize regional languages and along"
  },
  {
    "startMs": 1235600,
    "text": "with all the English and mixed language"
  },
  {
    "startMs": 1238880,
    "text": "uh uh documents and hence we'll soon uh"
  },
  {
    "startMs": 1242720,
    "text": "be able to train with the um uh feedback"
  },
  {
    "startMs": 1247039,
    "text": "we get from uh the current deployment."
  },
  {
    "startMs": 1249039,
    "text": "So yeah, we have started that uh"
  },
  {
    "startMs": 1250799,
    "text": "pipeline as well."
  },
  {
    "startMs": 1267439,
    "text": ">> [music]"
  }
]