Parsing the Epstein Files for JMail — Palak Agarwal & Omar Alhait, Reducto

Hello everyone, my Hello everyone, my name is Ple. I lead the name is Ple. I lead the developer relations department at Reductto. developer relations department at Reductto. Before we move on Before we move on Before we move on to Jmail, I'd like to to Jmail, I'd like to to Jmail, I'd like to tell you a little bit about what tell you a little bit about what tell you a little bit about what we do. we do. we do. We are an We are an We are an agent-based agent-based agent-based document platform, which document platform, which document platform, which essentially means essentially means essentially means we help we help we help teams teams teams transform transform transform unstructured unstructured unstructured documents into data documents into data documents into data that AI can accurately that AI can accurately that AI can accurately use. For use. For use. For context: over the past context: over the past context: over the past three years, we have raised $ three years, we have raised $ three years, we have raised $ 108 million from 108 million from 108 million from investors like investors like investors like ACCNC, Benchmark, and others, and ACCNC, Benchmark, and others, and ACCNC, Benchmark, and others, and processed over two processed over two processed over two billion documents billion documents billion documents for clients for clients for clients like Harvey, Scale, and one of the like Harvey, Scale, and one of the world's five largest hedge funds. Working world's five largest hedge funds. Working on this scale on this scale on this scale means we have means we have means we have come across every come across every come across every document you document you document you can can can imagine. We imagine. We imagine. We have improved the have improved the have improved the Reductto product so that it can Reductto product so that it can Reductto product so that it can handle anything. handle anything. Getting AI to work Getting AI to work with a pure PDF with a pure PDF with a pure PDF using a leading using a leading using a leading model isn't that model isn't that model isn't that difficult, but the truth is difficult, but the truth is difficult, but the truth is that most that most that most corporate corporate corporate documents are documents are documents are unstructured. unstructured. unstructured. This applies to This applies to This applies to scanned scanned scanned documents, documents, documents, rotated pages, and rotated pages, and rotated pages, and graphs. That's when graphs. That's when graphs. That's when things get really things get really things get really complicated. Reductto is the layer complicated. Reductto is the layer complicated. Reductto is the layer that sits directly that sits directly that sits directly in front of your model.

in front of your model. in front of your model. We handle We handle We handle complex layouts and complex layouts and complex layouts and also handle also handle also handle post-processing so you post-processing so you post-processing so you can focus can focus can focus on creating your on creating your on creating your product. The reason product. The reason product. The reason this works this works this works is because we is because we is because we use a use a rasterization-based approach, meaning rasterization-based approach, meaning we look at your we look at your we look at your document like a document like a document like a human would, instead of human would, instead of human would, instead of relying on relying on relying on embedded data. Everything embedded data. Everything embedded data. Everything works on the basis of our works on the basis of our works on the basis of our agent-based OCR agent-based OCR framework: instead of framework: instead of framework: instead of making one making one making one pass and giving a pass and giving a pass and giving a final answer, final answer, final answer, the model marks the model marks the model marks the points it is not the points it is not the points it is not sure about, corrects sure about, corrects sure about, corrects them and moves on to the them and moves on to the them and moves on to the next step. We next step. We next step. We will provide you with access to will provide you with access to will provide you with access to our studio. We our studio. We our studio. We made it made it made it stand-alone stand-alone stand-alone specifically for this specifically for this specifically for this session. In general, we have session. In general, we have session. In general, we have five main five main five main possibilities. Analysis, possibilities. Analysis, possibilities. Analysis, separation, separation, separation, extraction, extraction, extraction, editing and editing and editing and classification. Now, classification. Now, classification. Now, raise your hands. Who raise your hands. Who raise your hands. Who here has heard of Jmail or here has heard of Jmail or here has heard of Jmail or viewed Jmail before viewed Jmail before ? Good. Wow, that's quite a ? Good. Wow, that's quite a ? Good. Wow, that's quite a number of people. But number of people. But number of people. But I'll still briefly I'll still briefly I'll still briefly explain it if you're not explain it if you're not explain it if you're not aware. Jmail was aware. Jmail was aware. Jmail was created by Riley created by Riley created by Riley Walsh and Luke in Walsh and Luke in Walsh and Luke in five hours last five hours last five hours last November. This was November. This was November. This was when the when the Epstein files were released. The concept Epstein files were released. The concept is that it's is that it's is that it's essentially a fork of Gmail where essentially a fork of Gmail where essentially a fork of Gmail where you can log into your you can log into your you can log into your inbox like inbox like inbox like Epstein, and it's Epstein, and it's Epstein, and it's grown into a whole grown into a whole grown into a whole suite of services. You can suite of services. You can suite of services. You can imagine J-disk, J- imagine J-disk, J- flight, J-Amazon, J-photo and flight, J-Amazon, J-photo and flight, J-Amazon, J-photo and many more. By many more. By many more. By February, there were over 450 February, there were over 450 February, there were over 450 million visits and 18 million visits and 18 million visits and 18 million million million users users users accessing accessing accessing Jmail. And we are very Jmail. And we are very Jmail. And we are very proud that proud that proud that Reductto Reductto Reductto was responsible for was responsible for was responsible for analyzing all these analyzing all these analyzing all these documents. And that's a documents. And that's a documents. And that's a considerable number. That's considerable number. That's considerable number. That's over three million

over three million over three million pages of pages of pages of edited PDF edited PDF files, files, files, scanned scanned scanned documents, and documents, and documents, and messy government messy government messy government archives. Just so you archives. Just so you archives. Just so you understand, this understand, this understand, this included everything from included everything from included everything from very complex forms and very complex forms and very complex forms and tables to the tables to the tables to the illegible illegible illegible handwriting of many handwriting of many handwriting of many different people. And now different people. And now different people. And now I'll turn the floor over to Omar I'll turn the floor over to Omar , one of our , one of our founding engineers, who has founding engineers, who has been working closely been working closely been working closely with the Jail team since with the Jail team since with the Jail team since November. He November. He November. He will provide you with a will provide you with a will provide you with a technical overview and technical overview and technical overview and more. more. So, who's ready to So, who's ready to talk about talk about talk about documents? Hooray. documents? Hooray. OK. So, uh, yeah. Let OK. So, uh, yeah. Let 's go back, you know, 's go back, you know, 's go back, you know, to that fateful to that fateful to that fateful November night, shall we November night, shall we November night, shall we ? I was talking to the ? I was talking to the ? I was talking to the founding team founding team , and we were trying to , and we were trying to , and we were trying to figure out how figure out how figure out how to best use to best use to best use Reduct in this iconic Reduct in this iconic Reduct in this iconic cultural event that was going to be cultural event that was going to be presented to presented to the public, the public, the public, you know? It just so happens you know? It just so happens you know? It just so happens that we are really that we are really that we are really close with Luke, the close with Luke, the close with Luke, the founder of Kino, who founder of Kino, who founder of Kino, who works works works downstairs, and we were just downstairs, and we were just downstairs, and we were just discussing what would be discussing what would be discussing what would be cool and fun cool and fun cool and fun to create. Haha. And to create. Haha. And to create. Haha. And then the idea grew then the idea grew then the idea grew into something like: what into something like: what into something like: what if you could log if you could log if you could log into your email supposedly under into your email supposedly under into your email supposedly under Jeffrey Jeffrey Jeffrey Epstein's name and Epstein's name and Epstein's name and view it on view it on view it on his behalf? It was a his behalf? It was a his behalf? It was a really cool really cool really cool concept for a product, concept for a product, concept for a product, but there was one but there was one but there was one super-complex problem super-complex problem super-complex problem that actually became a that actually became a that actually became a barrier between us and the

barrier between us and the barrier between us and the finished solution. And finished solution. And finished solution. And you're looking at you're looking at you're looking at her part now, right? So I her part now, right? So I her part now, right? So I strongly advise strongly advise strongly advise you, if you have you, if you have you, if you have n't done so before, n't done so before, n't done so before, to send that to send that to send that cover page to cover page to cover page to ChatGPT and try to ChatGPT and try to ChatGPT and try to get some sort of get some sort of get some sort of result to result to result to see how see how see how accurate it is, you know? accurate it is, you know? This is a very complex This is a very complex problem, where the layouts problem, where the layouts problem, where the layouts are quite are quite are quite difficult to difficult to difficult to understand. The understand. The understand. The reading order here is also reading order here is also reading order here is also somewhat somewhat somewhat subjective, isn't subjective, isn't subjective, isn't it? So, we it? So, we it? So, we needed something needed something needed something to solve this to solve this to solve this problem and problem and problem and continue working continue working continue working on all these on all these on all these cool products. And this is where cool products. And this is where Redto came to the rescue. I want Redto came to the rescue. I want to show you what we to show you what we to show you what we used for used for used for a few of our a few of our a few of our products, okay? Oh, yeah, products, okay? Oh, yeah, products, okay? Oh, yeah, let me let me let me do this quickly. Well, okay do this quickly. Well, okay do this quickly. Well, okay . Oh. I'll . Oh. I'll . Oh. I'll drag it over now. Well, here it is.

drag it over now. Well, here it is. drag it over now. Well, here it is. So, you see a So, you see a So, you see a cool product here, right? J cool product here, right? J cool product here, right? J Flights. Actually, the whole Flights. Actually, the whole Flights. Actually, the whole story of story of story of Jeffrey Epstein and Jeffrey Epstein and Jeffrey Epstein and his associates' flights, right? his associates' flights, right? his associates' flights, right? He transported people by He transported people by He transported people by plane. He was plane. He was plane. He was transporting diplomats transporting diplomats . He himself flew to . He himself flew to . He himself flew to many places to his many places to his many places to his island and back, right? Uh island and back, right? Uh , the only thing standing between us , the only thing standing between us , the only thing standing between us and discovering all and discovering all and discovering all these flights was these these flights was these these flights was these rather poorly rather poorly rather poorly handwritten handwritten handwritten documents, right? Uh, documents, right? Uh, documents, right? Uh, again, if you again, if you again, if you thought that thought that thought that title page was title page was title page was a bit shoddy, a bit shoddy, a bit shoddy, think about the think about the think about the handwriting, right handwriting, right ? So here we see the ? So here we see the ? So here we see the pipeline that I pipeline that I pipeline that I actually set up actually set up actually set up back when, uh, we back when, uh, we back when, uh, we were trying to create a were trying to create a were trying to create a good configuration good configuration good configuration to make our to make our to make our parsing really parsing really parsing really good, right? And in good, right? And in good, right? And in this case, uh, this case, uh, this case, uh, for the for the for the flight scenario, we flight scenario, we flight scenario, we had an idea, what if had an idea, what if had an idea, what if we, uh, separate this we, uh, separate this we, uh, separate this into a separate product where you into a separate product where you into a separate product where you could just could just could just see all of his see all of his see all of his flights. Uh, and that flights. Uh, and that flights. Uh, and that meant taking all of meant taking all of meant taking all of his flight his flight his flight documentation and documentation and documentation and applying a applying a applying a separate separate separate data extraction configuration to it. So data extraction configuration to it. So data extraction configuration to it. So you'll see here, uh, you'll see here, uh, you'll see here, uh, a little tour a little tour a little tour of how Reduct works, of how Reduct works, of how Reduct works, right? Reduct in a right? Reduct in a document extraction scenario first document extraction scenario first parses the document, parses the document, parses the document, right? And this shows you right? And this shows you right? And this shows you that you can insert that you can insert that you can insert any any any document you document you document you want here and just want here and just want here and just hit "run," hit "run," hit "run," right? Customize right? Customize any of our features and any of our features and any of our features and see the results see the results see the results in real in real in real time, right? So that was

time, right? So that was time, right? So that was kind of a canvas kind of a canvas kind of a canvas for me to for me to for me to play with to get a sense of play with to get a sense of play with to get a sense of everything before everything before everything before scaling it up to a scaling it up to a scaling it up to a million different million different million different documents and quickly documents and quickly documents and quickly checking receipts checking receipts checking receipts in Reduct. Let us in Reduct. Let us in Reduct. Let us make sure that our make sure that our data extraction accuracy meets data extraction accuracy meets expectations. This is what expectations. This is what expectations. This is what we ended up with we ended up with we ended up with . You can . You can . You can see here in see here in see here in a few of these different a few of these different a few of these different documents some of the documents some of the documents some of the edits we edits we edits we had to work with. had to work with. had to work with. Woody Allen, I know Woody Allen, I know Woody Allen, I know this guy. But yeah, this guy. But yeah, this guy. But yeah, a lot of different a lot of different a lot of different scenarios that scenarios that scenarios that we had to we had to we had to deal with deal with deal with with one with one with one extraction scheme for extraction scheme for extraction scheme for flights, right? And we flights, right? And we analyzed all the analyzed all the documents very quickly. And by the way, a documents very quickly. And by the way, a documents very quickly. And by the way, a funny story about funny story about funny story about this, uh, is Alvin in this, uh, is Alvin in this, uh, is Alvin in the room? Alvin is in the room? Alvin is in the room? Alvin is in the room. So, uh, back the room. So, uh, back the room. So, uh, back when the when the when the first, uh, issue came out, I first, uh, issue came out, I first, uh, issue came out, I called Alvin, called Alvin, called Alvin, maybe a few times maybe a few times maybe a few times a night. Uh, he, he a night. Uh, he, he a night. Uh, he, he just happened to be just happened to be just happened to be on duty the on duty the on duty the night I was processing night I was processing night I was processing all these documents. And we all these documents. And we all these documents. And we tried to process tried to process tried to process them very, very quickly.

them very, very quickly. them very, very quickly. So, you know, we've So, you know, we've So, you know, we've put quite a strain on put quite a strain on put quite a strain on our infrastructure. our infrastructure. our infrastructure. And Alvin would send me And Alvin would send me And Alvin would send me a bunch of funny a bunch of funny a bunch of funny messages late messages late messages late at night, like, "Hey at night, like, "Hey at night, like, "Hey bro, can you bro, can you bro, can you slow down a bit?" Um, slow down a bit?" Um, slow down a bit?" Um, but yeah, we handled but yeah, we handled but yeah, we handled them. We dealt with them. We dealt with them. We dealt with them in time. And then, them in time. And then, them in time. And then, when we had this when we had this when we had this body of documents body of documents body of documents that we could, you that we could, you that we could, you know, take know, take know, take as the as the as the ultimate truth, ultimate truth, ultimate truth, we launched a billion we launched a billion we launched a billion different products, different products, different products, right? Paula right? Paula right? Paula touched on this a little bit, but touched on this a little bit, but touched on this a little bit, but I'll go back to I'll go back to I'll go back to Gmail. So, we've created Gmail. So, we've created Gmail. So, we've created everything from Jazm to Jlights everything from Jazm to Jlights everything from Jazm to Jlights to Gemini, right? Um, yes. to Gemini, right? Um, yes. to Gemini, right? Um, yes. Yes, and JCAL. This Yes, and JCAL. This Yes, and JCAL. This was actually a product was actually a product was actually a product I worked on I worked on I worked on myself. So there was a myself. So there was a myself. So there was a lot of interesting stuff there. Hmm, lot of interesting stuff there. Hmm, lot of interesting stuff there. Hmm, cool. So I think, cool. So I think, cool. So I think, you know, if you could you know, if you could you know, if you could take anything out of take anything out of take anything out of this, I actually this, I actually this, I actually want to say um, yeah, want to say um, yeah, want to say um, yeah, if you could if you could if you could take anything out of this, take anything out of this, take anything out of this, the simplicity here is the simplicity here is the simplicity here is to get a to get a to get a diagram, right?

diagram, right? diagram, right? Review Review Review the results, the results, the results, make sure they make sure they make sure they look like I look like I look like I expect, and then expect, and then expect, and then just deploy. just deploy. just deploy. Truth? So if I were to Truth? So if I were to Truth? So if I were to deploy this deploy this deploy this piece of code, I'll piece of code, I'll piece of code, I'll just redeploy just redeploy just redeploy it here for our it here for our it here for our use case use case , and I can get a , and I can get a , and I can get a direct API call or direct API call or direct API call or just a piece of just a piece of just a piece of pipeline code that I can pipeline code that I can call very quickly, right? Um, call very quickly, right? Um, so the point here is to so the point here is to so the point here is to upload your upload your upload your documents to this documents to this documents to this platform, platform, platform, start visualizing start visualizing start visualizing your data right away, and then your data right away, and then your data right away, and then instantly instantly instantly launch it into launch it into launch it into production. So I'll production. So I'll stop here quickly. Um, does stop here quickly. Um, does anyone have any questions at anyone have any questions at anyone have any questions at this point about Jmail this point about Jmail this point about Jmail or Reduct in general or Reduct in general or Reduct in general before we move on to a before we move on to a before we move on to a little little little interactive exercise interactive exercise ? ? Yes. Yes, great question. We great question. We can actually can actually can actually create this for you, create this for you, create this for you, right? Um, my right? Um, my right? Um, my funny analogy funny analogy funny analogy is that is that is that your extraction scheme your extraction scheme is like a brush on a is like a brush on a is like a brush on a Reduct canvas, right?

Reduct canvas, right? So she definitely So she definitely controls everything you controls everything you controls everything you get from the scheme. get from the scheme. So, for example, I So, for example, I can create a new can create a new can create a new pipeline here, right? and pipeline here, right? and pipeline here, right? and let's create a new let's create a new data mining pipeline, let's call it, I do data mining pipeline, let's call it, I do n't know, let's n't know, let's n't know, let's create a schema for create a schema for create a schema for resumes, because I want to resumes, because I want to resumes, because I want to parse a bunch of parse a bunch of parse a bunch of resumes, resumes, oh, sorry, oh, sorry, actually I'll do this actually I'll do this actually I'll do this really quickly, so I'll really quickly, so I'll really quickly, so I'll do this again, because it's do this again, because it's do this again, because it's really easy, I'll really easy, I'll really easy, I'll create a new create a new create a new mining pipeline here, by mining pipeline here, by mining pipeline here, by the way, we can the way, we can the way, we can do a lot of things with do a lot of things with do a lot of things with documents: documents: documents: split, split, split, extract, extract, extract, edit, and even edit, and even edit, and even classify. So classify. So classify. So what I'm going to do here is what I'm going to do here is create a schema for the create a schema for the resume. Oh. Uh, my Wi-Fi is resume. Oh. Uh, my Wi-Fi is resume. Oh. Uh, my Wi-Fi is n't helping much today n't helping much today n't helping much today . That's it. . That's it. Good. So, I have an Good. So, I have an interesting scheme. And now interesting scheme. And now interesting scheme. And now I have to decide whether I have to decide whether I have to decide whether this scheme suits me this scheme suits me this scheme suits me or not. Yes. She will or not. Yes. She will or not. Yes. She will truly give me truly give me truly give me exactly what I exactly what I exactly what I ask for. And you know, once ask for. And you know, once ask for. And you know, once I upload a I upload a I upload a document here, I can document here, I can document here, I can realize, "ah, this is realize, "ah, this is realize, "ah, this is n't really what n't really what n't really what I need," I need," I need," right? So I have a right? So I have a right? So I have a bunch of handy resumes.

bunch of handy resumes. bunch of handy resumes. I'm going to paste I'm going to paste I'm going to paste them in here, download them, them in here, download them, and then just and then just run this. So yes, I run this. So yes, I run this. So yes, I hope I hope I hope I answered your answered your answered your question. Are there any more question. Are there any more question. Are there any more questions while we questions while we questions while we wait? wait? How do you perform How do you perform validation? Are validation? Are you talking about the you talking about the download status? It's download status? It's download status? It's just such an interesting just such an interesting just such an interesting loading state, loading state, loading state, after all, right? Our after all, right? Our after all, right? Our product is a set of product is a set of product is a set of different models and different models and different models and different VLMs that we different VLMs that we different VLMs that we deploy, right? Uh, deploy, right? Uh, deploy, right? Uh, I...I... I...I... I...I... sorry, it's not about sorry, it's not about sorry, it's not about loading. I loading. I loading. I followed what followed what followed what you were saying. you were saying. Yes. Oh, that's a great Oh, that's a great question. So, if question. So, if question. So, if I understand correctly, I understand correctly, I understand correctly, what exactly do you what exactly do you what exactly do you mean?

mean? mean? Yes. Yes. This particular Yes. Yes. This particular Yes. Yes. This particular schema generation schema generation schema generation I just used I just used I just used works out of the box, works out of the box, works out of the box, it's just a quick it's just a quick it's just a quick query to LLM that query to LLM that query to LLM that builds a JSON schema exactly builds a JSON schema exactly builds a JSON schema exactly according to my query. according to my query. However, we have However, we have advanced advanced advanced schema generation, right? And it schema generation, right? And it schema generation, right? And it works like this: works like this: works like this: let's say I let's say I let's say I parsed my parsed my parsed my documents, it documents, it documents, it will take the results from will take the results from will take the results from these documents and these documents and these documents and adapt the schema description adapt the schema description adapt the schema description specifically for them. specifically for them. So if you're looking for, So if you're looking for, for example, something very for example, something very for example, something very fast to just fast to just fast to just output to the screen and output to the screen and output to the screen and see the results, see the results, see the results, I would recommend I would recommend I would recommend regular regular regular schema generation. But if you are schema generation. But if you are schema generation. But if you are more interested in more interested in more interested in the nuances, I definitely the nuances, I definitely the nuances, I definitely recommend recommend recommend parsing the documents first parsing the documents first , and then transferring them , and then transferring them , and then transferring them to the system. All you to the system. All you to the system. All you need to do is need to do is need to do is simply click " simply click " improve improve improve schema creation". And we can, you schema creation". And we can, you schema creation". And we can, you know, first know, first know, first analyze the analyze the analyze the documents, documents, documents, try to understand try to understand what they contain, what they contain, what they contain, so that we can so that we can so that we can adapt the adapt the adapt the schema descriptions directly schema descriptions directly schema descriptions directly to them. Of course?

to them. Of course? to them. Of course? Class. Class. Class. Class. Class. Class. Class. Class. you had some you had some editing. editing. editing. Yes. Yes. Right. Yes. Well, I Right. Yes. Well, I mean, that's a mean, that's a mean, that's a great question, great question, great question, right? Reduct also right? Reduct also right? Reduct also parses PDF metadata, parses PDF metadata, parses PDF metadata, right? So sometimes, and right? So sometimes, and right? So sometimes, and that's honestly that's honestly that's honestly indicative of who indicative of who indicative of who was doing was doing was doing the editing, the editing, the editing, right? Sometimes right? Sometimes right? Sometimes the editing was the editing was the editing was n't really that n't really that n't really that good, was it? And as good, was it? And as good, was it? And as a result, a result, a result, the responsibility the responsibility the responsibility fell on us, the fell on us, the fell on us, the JML team, we had to JML team, we had to JML team, we had to do a bunch of do a bunch of do a bunch of editing, right? Because you editing, right? Because you editing, right? Because you can't just can't just can't just stick a rectangle stick a rectangle stick a rectangle on a PDF and expect on a PDF and expect on a PDF and expect it to be fully it to be fully it to be fully editable, right? editable, right? editable, right? Um, yes, it seems there was Um, yes, it seems there was Um, yes, it seems there was some misunderstanding on the some misunderstanding on the some misunderstanding on the part of whoever part of whoever part of whoever edited it. Hmm, I'll say edited it. Hmm, I'll say that for the files that for the files that for the files we received, when I we received, when I we received, when I reviewed them myself, that's reviewed them myself, that's reviewed them myself, that's not the case, right? Some not the case, right? Some not the case, right? Some files were files were files were edited edited edited correctly, but there were also correctly, but there were also correctly, but there were also some where we could some where we could some where we could see who was behind the see who was behind the see who was behind the edit, and edit, and edit, and that became a very that became a very that became a very serious situation, serious situation, serious situation, right? Because we're right? Because we're right? Because we're dealing with people's data dealing with people's data , and our data , and our data , and our data is edited for a reason is edited for a reason is edited for a reason , right?

, right? And you want to make And you want to make sure that you're not the one who ends up being sure that you're not the one who ends up being sued sued at the end of the day because of it. at the end of the day because of it. at the end of the day because of it. So yes, there So yes, there So yes, there was a was a was a huge burden of huge burden of huge burden of responsibility on our shoulders responsibility on our shoulders responsibility on our shoulders to make sure that to make sure that to make sure that everything that was supposed to be everything that was supposed to be everything that was supposed to be edited was edited was edited was actually edited actually edited . Our tool . Our tool . Our tool worked too worked too worked too well. well. Class. Yes. first me, Class. Yes. first me, then you then you questions or thousands of questions or thousands of different things. different things. Yes. Yes. Yes. That's a Yes. Yes. Yes. That's a great question. Um, great question. Um, I'll give you an example, I'll give you an example, okay? There was a diagram okay? There was a diagram okay? There was a diagram I looked at, maybe on I looked at, maybe on I looked at, maybe on Friday, it had Friday, it had Friday, it had about 156 fields, right? about 156 fields, right? and all these different fields and all these different fields needed something needed something needed something completely different, completely different, completely different, you know? Therefore, I you know? Therefore, I you know? Therefore, I urge you urge you urge you to test our to test our to test our scheme. I think our, uh scheme. I think our, uh , , , schema nesting restrictions schema nesting restrictions schema nesting restrictions are somewhere around are somewhere around are somewhere around four or five four or five four or five levels. But if you levels. But if you levels. But if you want to, you know, want to, you know, want to, you know, ask about everything in the ask about everything in the ask about everything in the world, technically you world, technically you world, technically you can still fit can still fit can still fit it into the diagram.

it into the diagram. Do you have entity recognition? Entity recognition. Can you Entity recognition. Can you give an example? give an example? That's a good question. That's a good question. Um, we end up Um, we end up Um, we end up doing a lot of doing a lot of doing a lot of post-processing, post-processing, post-processing, right? Let's say at the right? Let's say at the right? Let's say at the data extraction stage. data extraction stage. data extraction stage. Personally, I'm a little further Personally, I'm a little further Personally, I'm a little further removed from, removed from, removed from, say, the ML component. say, the ML component. say, the ML component. But I can say that But I can say that But I can say that we do a lot of we do a lot of we do a lot of post-processing to post-processing to post-processing to essentially connect essentially connect essentially connect entities together. entities together. entities together. For example, we have a For example, we have a For example, we have a citation model, right citation model, right ? Where, you know, one ? Where, you know, one ? Where, you know, one model can be model can be model can be responsible for responsible for responsible for the extraction, right? But the extraction, right? But the extraction, right? But then there could be another then there could be another then there could be another model that, after model that, after model that, after data extraction, is data extraction, is data extraction, is responsible for responsible for responsible for where exactly where exactly where exactly each fragment came from, right? each fragment came from, right? Nvidia, by the way, Nvidia, by the way, released an interesting released an interesting released an interesting article about this, which article about this, which article about this, which I only learned about yesterday I only learned about yesterday . I think it's . I think it's . I think it's called "find called "find called "find everything" or something like that. Where everything" or something like that. Where everything" or something like that. Where they're trying they're trying they're trying to solve the to solve the to solve the bottleneck problem, bottleneck problem, bottleneck problem, you know, defining you know, defining you know, defining citations and the speed with citations and the speed with citations and the speed with which you can find a which you can find a which you can find a citation in a document.

citation in a document. If you are interested, If you are interested, we are hiring, we are hiring, we are hiring, so come see us when you are so come see us when you are so come see us when you are finished. finished. Sorry. Sorry. Question. Question. Yes, of course. So, I'll Yes, of course. So, I'll try to check try to check try to check if there is an image in if there is an image in if there is an image in one of the resumes I one of the resumes I one of the resumes I uploaded. uploaded. Let's take a look. Let's take a look. Yes. So we have Yes. So we have Yes. So we have certain models for certain models for certain models for shapes that define shapes that define ...let me ...let me ...let me quickly restart quickly restart quickly restart this so it can this so it can this so it can return shapes. return shapes. Ah, that's a great Ah, that's a great question. Alvin, did you question. Alvin, did you question. Alvin, did you hear that? hear that? hear that? Sorry. Sorry. Sorry. No, no. I will explain.

No, no. I will explain. No, no. I will explain. The reason I The reason I The reason I asked Alvin is asked Alvin is asked Alvin is because he because he because he was working on our was working on our was working on our hybrid OCR model, hybrid OCR model, hybrid OCR model, right? We have a right? We have a right? We have a special OCR special OCR configuration that you configuration that you configuration that you can enable here, can enable here, can enable here, and it and it and it will intelligently determine what will intelligently determine what will intelligently determine what to prioritize. In to prioritize. In to prioritize. In some cases, we some cases, we some cases, we select purely OCR select purely OCR data or metadata from a data or metadata from a data or metadata from a PDF file. Yes, we PDF file. Yes, we PDF file. Yes, we did a lot of did a lot of did a lot of work here. And this is not an easy work here. And this is not an easy work here. And this is not an easy decision, so to decision, so to decision, so to speak, not a simple speak, not a simple speak, not a simple answer. This is a topic answer. This is a topic answer. This is a topic for a whole separate for a whole separate for a whole separate report. report. Yes. I'll be honest, Yes. I'll be honest, that's exactly what that's exactly what that's exactly what this conveyor belt does, this conveyor belt does, this conveyor belt does, you know? So, you know? So, you know? So, first we first we first we analyze, yeah, analyze, yeah, analyze, yeah, wait, let wait, let wait, let me get the numbers back.

me get the numbers back. Steeply. So while I'm Steeply. So while I'm restarting this, restarting this, restarting this, let me let me let me pick another example pick another example pick another example of what we're doing here of what we're doing here . And, in fact, we . And, in fact, we . And, in fact, we will have to will have to will have to wait now. But wait now. But wait now. But in reality, during in reality, during in reality, during data mining, we data mining, we data mining, we take the results of take the results of take the results of the analysis and pass the analysis and pass the analysis and pass them into some further them into some further them into some further processing pipeline. So processing pipeline. So processing pipeline. So analysis is, you could analysis is, you could analysis is, you could say, what say, what say, what made Reducto made Reducto made Reducto Reducto from the Reducto from the Reducto from the beginning, beginning, beginning, right? And then we can right? And then we can right? And then we can do a lot of cool do a lot of cool do a lot of cool things after things after things after Reducto works. So once we Reducto works. So once we Reducto works. So once we have the have the content analyzed, we can, content analyzed, we can, for example, break for example, break for example, break the document down into different the document down into different the document down into different areas and figure out, areas and figure out, areas and figure out, if I'm analyzing a if I'm analyzing a math textbook, where exactly each each section begins and ends. And then for section begins and ends. And then for section begins and ends. And then for each section, I each section, I each section, I can do can do can do data mining data mining data mining based on that to based on that to based on that to get a summary of the get a summary of the get a summary of the section or something section or something section or something like that. So analysis like that. So analysis is really the foundation is really the foundation is really the foundation for all these other for all these other document processing agent functions.

Dude, if I Dude, if I answer this answer this answer this question, I question, I question, I could lose my job tomorrow. could lose my job tomorrow. could lose my job tomorrow. So, I mean, this is So, I mean, this is ... Yeah, go ahead. ... Yeah, go ahead. I think it also I think it also depends on depends on depends on what data we collect. what data we collect. So we specifically So we specifically train on certain train on certain train on certain types of data. If we types of data. If we types of data. If we see that the model does not see that the model does not see that the model does not work well with work well with work well with certain types of data. certain types of data. certain types of data. For example, we see that For example, we see that For example, we see that remembering remembering remembering long context is not long context is not long context is not very good. Then we very good. Then we very good. Then we should try to should try to should try to fix it. Specifically fix it. Specifically fix it. Specifically for extraction, it's for extraction, it's for extraction, it's not really like a needle not really like a needle not really like a needle in a haystack. It's like in a haystack. It's like in a haystack. It's like all the needles in a haystack all the needles in a haystack all the needles in a haystack , if , if , if you think about it. So this is a you think about it. So this is a you think about it. So this is a really complex really complex really complex research research research problem that our problem that our machine learning specialists are machine learning specialists are somehow somehow somehow doing well. But yes, doing well. But yes, doing well. But yes, I'd be happy I'd be happy I'd be happy to talk more to talk more to talk more later. later. Yes. Yes. Oh, we neglected you a bit Oh, we neglected you a bit Oh, we neglected you a bit . Yes. How are . Yes. How are . Yes. How are you? You will... you? You will... Yes, you should Yes, you should Yes, you should follow our follow our follow our social media social media , and you will find out about , and you will find out about , and you will find out about it very soon. I think it very soon. I think performance testing is performance testing is always a topic for always a topic for always a topic for discussion and something discussion and something discussion and something we want to we want to we want to focus on, focus on, focus on, especially because especially because especially because someone who wants to someone who wants to someone who wants to do document extraction do document extraction might not know where might not know where to turn or who is to turn or who is to turn or who is best without best without best without testing it out for testing it out for testing it out for themselves. So themselves. So themselves. So testing is something testing is something testing is something we're paying we're paying we're paying a lot of attention to, a lot of attention to, a lot of attention to, especially in the coming especially in the coming especially in the coming months, to show months, to show months, to show people how people how people how effective we are. You know, we're effective we are. You know, we're effective we are. You know, we're very confident in our very confident in our very confident in our models and that models and that models and that we're the leading we're the leading document data extraction solution in the market.

But we need more But we need more data points that we data points that we data points that we can rely on and can rely on and can rely on and that the entire industry can that the entire industry can that the entire industry can take for granted, take for granted, take for granted, without having to run our own without having to run our own without having to run our own tests and tests and tests and run the run the run the entire entire entire ARC AGI dataset through them, to understand ARC AGI dataset through them, to understand ARC AGI dataset through them, to understand how Reducto is better how Reducto is better how Reducto is better than other solutions. than other solutions. Yes. So we will Yes. So we will publish publish publish the results of our the results of our the results of our tests soon. But the real tests soon. But the real tests soon. But the real answer is answer is answer is that, unfortunately, that, unfortunately, that, unfortunately, you usually you usually you usually have to have to have to create your own create your own create your own tests. Therefore, almost tests. Therefore, almost tests. Therefore, almost every one of our clients every one of our clients every one of our clients usually has their usually has their usually has their own benchmark. own benchmark. own benchmark. They say, "Oh, we They say, "Oh, we They say, "Oh, we evaluate evaluate document processing service providers." We are document processing service providers." We are going to do going to do going to do direct comparisons of Reducto direct comparisons of Reducto direct comparisons of Reducto with other solutions, and with other solutions, and with other solutions, and sometimes they sometimes they sometimes they use our use our use our APIs not entirely APIs not entirely APIs not entirely efficiently due to efficiently due to efficiently due to incorrect incorrect incorrect settings. So we settings. So we settings. So we work closely with work closely with work closely with them, tell them what them, tell them what them, tell them what to enable, what the to enable, what the to enable, what the results will be, results will be, results will be, let them let them let them decide, and often decide, and often decide, and often we win these we win these performance comparisons. So, performance comparisons. So, as with most as with most as with most projects, collect a projects, collect a projects, collect a simple set of data. simple set of data. simple set of data. You can make a You can make a You can make a quick assessment “by eye quick assessment “by eye ”: just look ”: just look ”: just look in the studio to see if the in the studio to see if the in the studio to see if the raw data makes sense. And raw data makes sense. And raw data makes sense. And then when you want then when you want then when you want to scale, to scale, to scale, take 100 take 100 take 100 documents, a thousand documents, a thousand documents, a thousand documents and do documents and do documents and do something more something more something more systematic. You systematic. You systematic. You can also can also can also work with our work with our work with our team, we have team, we have team, we have many specialists who many specialists who many specialists who help clients help clients help clients conduct assessments conduct assessments , yes.

, yes. , yes. Steeply. Are there any more Steeply. Are there any more Steeply. Are there any more questions? Yes, questions? Yes, questions? Yes, of course. Well, yes, of course. Well, yes, of course. Well, yes, absolutely. Hmm, I absolutely. Hmm, I absolutely. Hmm, I think so. I'll answer think so. I'll answer think so. I'll answer this fairly honestly, this fairly honestly, this fairly honestly, okay? I think when I okay? I think when I okay? I think when I was working on this was working on this was working on this solution, or more precisely solution, or more precisely solution, or more precisely this problem, I this problem, I this problem, I was trying to figure out was trying to figure out was trying to figure out how best how best how best to turn our to turn our to turn our data parsing into a data parsing into a data parsing into a product. To be honest product. To be honest , at first I made a , at first I made a , at first I made a pretty primitive pretty primitive pretty primitive thing—something like a thing—something like a thing—something like a chatbot with chatbot with chatbot with RAG technology, where you could RAG technology, where you could RAG technology, where you could ask ask ask anything about anything about anything about emails. And emails. And emails. And this could be very, this could be very, this could be very, very useful for very useful for very useful for some journalist. some journalist. However, I think it However, I think it speaks to the hundreds of speaks to the hundreds of speaks to the hundreds of millions of people who millions of people who millions of people who have used Gmail, and the have used Gmail, and the have used Gmail, and the tens of tens of tens of millions who millions who millions who continue to log in continue to log in continue to log in every week. This every week. This every week. This confirms the fact confirms the fact confirms the fact that a quality product that a quality product that a quality product on top of a great and on top of a great and on top of a great and innovative innovative innovative solution for solution for solution for extracting data from extracting data from extracting data from documents is exactly what documents is exactly what documents is exactly what will work as a will work as a will work as a double whammy.

double whammy. double whammy. Yes. Would you Yes. Would you Yes. Would you read three million read three million read three million scanned scanned scanned documents? Honestly, no. documents? Honestly, no. documents? Honestly, no. And frankly, And frankly, And frankly, when I saw the when I saw the when I saw the volume, I thought, "This is volume, I thought, "This is volume, I thought, "This is too much too much too much work for me." But work for me." But work for me." But once Omar once Omar once Omar worked it out, I just worked it out, I just worked it out, I just went in and started looking up went in and started looking up went in and started looking up random famous random famous random famous people for fun. That's the people for fun. That's the beauty of it, right? It beauty of it, right? It seems to me that seems to me that seems to me that unstructured data unstructured data unstructured data is everywhere. And sometimes all you is everywhere. And sometimes all you is everywhere. And sometimes all you need to do need to do is remove them and is remove them and is remove them and then do whatever you want with them then do whatever you want with them then do whatever you want with them . This is, . This is, . This is, in fact, what in fact, what all our all our clients do conceptually. So I think clients do conceptually. So I think clients do conceptually. So I think Gmail is a pretty Gmail is a pretty Gmail is a pretty interesting example, but interesting example, but interesting example, but it's actually what's it's actually what's it's actually what's happening in the happening in the happening in the industry. industry. industry. Yes. Did you see this? Yes. Did you see this? Yes. Did you see this? Come here. It's Come here. It's Come here. It's just hard work, just hard work, just hard work, right? You know, we right? You know, we right? You know, we have a data markup team have a data markup team have a data markup team , right? We have to , right? We have to sift through a huge sift through a huge number of documents number of documents if we want to if we want to if we want to train a new model on them train a new model on them train a new model on them . And to be very, . And to be very, very confident in very confident in very confident in the results or in the the results or in the the results or in the integrity of this data, integrity of this data, integrity of this data, you know? So the you know? So the you know? So the short answer short answer short answer for you is: it's a lot of for you is: it's a lot of for you is: it's a lot of hard, slow, and hard, slow, and hard, slow, and exhausting work, exhausting work, exhausting work, right? But ultimately, that's what right? But ultimately, that's what right? But ultimately, that's what forms the forms the forms the true beauty of a true beauty of a true beauty of a machine machine machine learning model, isn't it?

learning model, isn't it? In fact, nothing In fact, nothing can replace that slow, can replace that slow, can replace that slow, very hard work of very hard work of very hard work of classification, which classification, which classification, which then becomes the basis then becomes the basis then becomes the basis for training the model. for training the model. for training the model. And then you can And then you can And then you can figure out how to figure out how to figure out how to distribute your distribute your distribute your data and how to actually data and how to actually data and how to actually train train train on your on your on your dataset. Ultimately dataset. Ultimately , at the end of the day, you , at the end of the day, you , at the end of the day, you still have to still have to still have to mark up this data. mark up this data. mark up this data. Class. I think we'll Class. I think we'll Class. I think we'll take, yeah, we take, yeah, we take, yeah, we could take another one, could take another one, could take another one, but, um, yeah. Yes. So, but, um, yeah. Yes. So, but, um, yeah. Yes. So, to repeat to repeat to repeat the question, let me the question, let me the question, let me know how know how know how well I understood it well I understood it , okay? But it seems , okay? But it seems that there are cases where that there are cases where that there are cases where you really need to you really need to you really need to customize the logic customize the logic customize the logic for different documents. for different documents. Is it about Is it about specific models, specific models, specific models, right? Sometimes the metadata in right? Sometimes the metadata in right? Sometimes the metadata in certain documents certain documents certain documents may be may be may be structured in a structured in a structured in a specific way, and specific way, and specific way, and you want to take that into account you want to take that into account you want to take that into account for specific for specific use cases, right? That's a use cases, right? That's a great question, great question, great question, really. And this is where really. And this is where our agent approach comes in handy, our agent approach comes in handy, right? You can right? You can right? You can apply all these apply all these apply all these different configurations different configurations different configurations and system prompts and system prompts and system prompts to individual models, to individual models, to individual models, which will perform which will perform which will perform processing specifically processing specifically processing specifically tailored to your tailored to your tailored to your requests. For example, requests. For example, requests. For example, if you need to if you need to if you need to do complex do complex do complex shape extraction: shape extraction: shape extraction: let's say I have a bunch of let's say I have a bunch of let's say I have a bunch of logos and I want logos and I want logos and I want to determine what to determine what to determine what distinguishes, say, distinguishes, say, distinguishes, say, Lincoln Financial from Google, and Lincoln Financial from Google, and Lincoln Financial from Google, and how to separate those how to separate those how to separate those logos, right? I could

logos, right? I could logos, right? I could add more add more add more detail to our detail to our detail to our prompts in this prompts in this prompts in this configuration to configuration to configuration to tailor tailor tailor the results to the results to the results to my specific needs. So if my specific needs. So if my specific needs. So if you need you need you need specific metadata, specific metadata, specific metadata, say, some layout say, some layout say, some layout might be very hard might be very hard might be very hard to read, to read, to read, like in a like in a like in a construction construction construction drawing, right? I would drawing, right? I would drawing, right? I would spend a lot of time spend a lot of time spend a lot of time tweaking the tweaking the tweaking the layout hint layout hint layout hint to understand to understand to understand how well my how well my how well my current current current extraction solution works with extraction solution works with extraction solution works with all the different all the different all the different construction construction construction drawings, I think, drawings, I think, drawings, I think, yeah, great. Does yeah, great. Does yeah, great. Does this answer your this answer your this answer your question? question? question? Fantasy. OK. Fantasy. OK. Now we want to Now we want to turn this into an turn this into an turn this into an interactive exercise. interactive exercise. Yes. So, um, if Yes. So, um, if you want, we'll basically you want, we'll basically you want, we'll basically give you, you know, um, give you, you know, um, give you, you know, um, as much time as as much time as as much time as you want. I think we have you want. I think we have about 30 minutes left. Um, about 30 minutes left. Um, you can you can you can sign up sign up sign up today at, uh, the today at, uh, the today at, uh, the studio and basically studio and basically studio and basically just start, uh, just start, uh, just start, uh, working on a working on a working on a solution for solution for solution for extracting data from extracting data from extracting data from documents, right? So what documents, right? So what documents, right? So what you'll see here you'll see here you'll see here in the QR code is something that you in the QR code is something that you in the QR code is something that you can just can just can just paste paste paste directly into a directly into a directly into a cloud code or a codec cloud code or a codec cloud code or a codec or whatever and or whatever and or whatever and immediately start immediately start immediately start working on working on working on extracting data from extracting data from extracting data from documents. Uh, so documents. Uh, so documents. Uh, so we'll give you about 30 we'll give you about 30 we'll give you about 30 minutes, and somewhere around, minutes, and somewhere around, minutes, and somewhere around, you know, maybe you know, maybe you know, maybe five minutes before the five minutes before the five minutes before the end, we'd like to end, we'd like to end, we'd like to see what you've see what you've see what you've got there. We've got a got there. We've got a got there. We've got a bunch of cool merch and bunch of cool merch and bunch of cool merch and some great some great some great prizes for you to prizes for you to prizes for you to just show us

just show us just show us what you can, um, do in what you can, um, do in what you can, um, do in terms of extracting terms of extracting terms of extracting data from documents. data from documents. data from documents. Um, this might Um, this might Um, this might just be an idea, um, and just be an idea, um, and just be an idea, um, and some kind of one-time some kind of one-time some kind of one-time request, but, uh, don't be request, but, uh, don't be request, but, uh, don't be shy, just give it a shy, just give it a shy, just give it a try. That try. That try. That sounds good? sounds good? sounds good? Plus, we'll be there to help Plus, we'll be there to help Plus, we'll be there to help if if if you have any you have any you have any questions about questions about questions about how to use how to use how to use Redto. Uh, I want to Redto. Uh, I want to Redto. Uh, I want to work for you, work for you, work for you, you know, I want you know, I want you know, I want to learn about your to learn about your to learn about your models, how you models, how you models, how you train them, or train them, or train them, or something. Um, we'd be something. Um, we'd be something. Um, we'd be happy to answer. That happy to answer. That happy to answer. That sounds good?