WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=dZgA6btnb_s NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/dZgA6btnb_s.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:01.309 --> 00:00:03.309 [music] s2 00:00:12.600 --> 00:00:13.400 Hey, thanks for coming. s3 00:00:13.400 --> 00:00:14.680 My name is Vinod Ganesh. s4 00:00:14.680 --> 00:00:16.520 I'm the CEO of Kepler. s5 00:00:16.520 --> 00:00:18.960 We have a few of my teammates here as well. s6 00:00:18.960 --> 00:00:21.080 I'm excited to be here to chat with you. s7 00:00:21.080 --> 00:00:27.640 So, the initial inception of this session comes from the World's Fair for AIE. s8 00:00:27.640 --> 00:00:33.040 Where we discussed at a high level Kepler as a whole and how we do what we do. s9 00:00:33.040 --> 00:00:44.880 Um and our big claim to fame right now and there's a new bunch of other stuff is Anthropic did a full deep dive into our architecture, our technology, and the biggest question that I got after the World's Fair presentation s10 00:00:44.880 --> 00:00:47.400 was well, how do you actually do this? s11 00:00:47.400 --> 00:00:56.480 Um so, our partners and friends at Turbo puffer have given me this session to talk about how we use Turbo puffer as a component of this overarching ecosystem. s12 00:00:56.480 --> 00:01:01.560 Um also shameless plug, Jack and I are doing a talk on the main stage in about 25 minutes. s13 00:01:01.560 --> 00:01:11.040 So, we're going to power through this and if you want to hear about fundamental investing in the age of AI, uh feel free to come to that 25 minutes from now. s14 00:01:11.040 --> 00:01:14.120 Um so, to talk a little bit about uh so, quick background of myself. s15 00:01:14.120 --> 00:01:14.880 Um again, I'm Vinod. s16 00:01:14.880 --> 00:01:16.560 I'm the CEO of Kepler. s17 00:01:16.560 --> 00:01:18.160 Uh began my career at Palantir. s18 00:01:18.160 --> 00:01:20.160 I led the compute layer there. s19 00:01:20.160 --> 00:01:25.760 Um then built my own data as a service company and then was head of business engineering at Citadel. s20 00:01:25.760 --> 00:01:40.360 Um so, a lot of what we learned here in building uh Kepler comes directly from the experiences of understanding how to build and harness or a substrate or whatever the cool new word is in pretty aggressive environments s21 00:01:40.360 --> 00:01:44.720 where accuracy and fidelity matter pretty aggressively. s22 00:01:44.720 --> 00:01:47.560 So, let's talk a little bit about what Kepler does. s23 00:01:47.560 --> 00:01:49.880 Um I know we're um we've been in a lot of places. s24 00:01:49.880 --> 00:01:54.120 We're excited to be part of Gemini Force um blog post today as well. s25 00:01:54.120 --> 00:01:55.880 Uh they've a new model out. s26 00:01:55.880 --> 00:02:08.479 Um so, the core tenant is um, an analyst, if you're actually working in finance, you never show a client or your PM a number that they can't explicitly validate or trace. s27 00:02:08.479 --> 00:02:18.080 So, when you see something like revenue grew 12.4% to 418.2 billion, these are actual numbers that need to be correct. s28 00:02:18.080 --> 00:02:22.280 The challenge is we're all operating off of probability machines. s29 00:02:22.280 --> 00:02:24.600 LLMs are at their core probability machines. s30 00:02:24.600 --> 00:02:32.640 The same thing that makes them so creative and cool is the same thing that makes them very difficult to use in mission-critical situations. s31 00:02:32.640 --> 00:02:36.640 So, a natural thing you see is people that attach sources after the fact. s32 00:02:36.640 --> 00:02:42.360 Now, this is cool, but here's the problem that happens when you just attach a source after the fact. s33 00:02:42.360 --> 00:02:47.920 First, LLMs don't actually read your documents, despite what many people think. s34 00:02:47.920 --> 00:02:56.360 What they do is they chunk a bunch of things together, they do a semantic search, and they retrieve information that may be semi-relevant, if you're lucky. s35 00:02:56.360 --> 00:03:00.519 So, the first big challenge a lot of people have is we have the wrong source. s36 00:03:00.519 --> 00:03:08.720 Like, we've pulled information from Seeking Alpha or Venu's random YOLO blog that says Palantir will be worth $100,000 a share. s37 00:03:08.720 --> 00:03:12.000 Or, we have the right source and the wrong number. s38 00:03:12.000 --> 00:03:18.840 For example, when you read a 10K or a 10Q, you can imagine a world where I see the number 17. s39 00:03:18.840 --> 00:03:20.840 Well, what does that correspond to? s40 00:03:20.840 --> 00:03:26.959 Is it 17 million, 17 billion, is it 17 million yen, 17 million dollars? s41 00:03:26.959 --> 00:03:34.560 Is it 17 million was the revenue of this company in the year 2023 in the 10K that's in 2025? s42 00:03:34.560 --> 00:03:36.800 Or, is 17 the page number? s43 00:03:36.800 --> 00:03:46.320 Like, all of these things are considerations that occur when we're actually working on any kind of environment that requires verifiability or numerical accuracy. s44 00:03:46.320 --> 00:03:48.680 Finally, there's a piece of mental math. s45 00:03:48.680 --> 00:04:00.160 Growth rates, margins, multiples, models are not actually doing Now, in some cases, they write code that does do math, but models themselves are not actually doing math in a way that ensures s46 00:04:00.160 --> 00:04:02.480 fidelity and accuracy. s47 00:04:02.480 --> 00:04:10.280 So, what we end up having to develop is an architecture where the model sites the numbers, but never writes them. s48 00:04:10.280 --> 00:04:17.600 At its core and the magic that powers Kepler is retrieval finds the right source, thank you TurboPuffer, which we'll talk about. s49 00:04:17.600 --> 00:04:25.960 Code does the math, and code meaning the model is not writing the math, but predetermined a deterministic code that we humans have written. s50 00:04:25.960 --> 00:04:28.440 And we have a post-processing filter. s51 00:04:28.440 --> 00:04:30.080 So, let's break this down. s52 00:04:30.080 --> 00:04:35.680 So, what we did is we processed the entire corpus of SEC data. s53 00:04:35.680 --> 00:04:38.720 Now, that's 30 plus million documents. s54 00:04:38.720 --> 00:04:44.880 Um our friends at TurboPuffer will be hopefully working with us fairly soon on how we actually surface this to all of you. s55 00:04:44.880 --> 00:04:50.760 Um but we ran a pre-processing step, and we registered what we're calling citations. s56 00:04:50.760 --> 00:04:57.280 And so, citations are effectively the deterministic numerical numbers that we have pulled out of filings. s57 00:04:57.280 --> 00:05:08.400 Things like revenue numbers, things like an income statement, a balance sheet, things like MD&A that actually says, "Hey, my revenue grew 37% in human-written text." s58 00:05:08.400 --> 00:05:11.160 There's about 20,000 of those. s59 00:05:11.160 --> 00:05:26.600 There's three individual filing facts that we've revealed as citable anchors, which effectively means our ability to tell you where a number came from comes down to the ability to be able to click on a number and actually reference back a core document where that came from. s60 00:05:26.600 --> 00:05:32.520 When we're running in production, we actually refuse citations that we can't verify. s61 00:05:32.520 --> 00:05:39.680 Meaning, if I have a number that's not cited, the model is structurally incapable of producing this output. s62 00:05:39.720 --> 00:05:44.840 Finally, uh the number of citations that the filter has correctly extracted, we got about 21,000. s63 00:05:44.840 --> 00:05:48.080 I'll I'll through this in a little bit more detail. s64 00:05:48.360 --> 00:05:57.120 How Kepler works and the piece and anthropic, if you want more information Google Kepler anthropic, you can read about is the language models themselves are planning. s65 00:05:57.120 --> 00:06:03.440 Deterministic software handles every number or even more generally every single fact. s66 00:06:03.440 --> 00:06:04.880 So, let's break this down. s67 00:06:04.880 --> 00:06:06.440 A user asks the question. s68 00:06:06.440 --> 00:06:15.920 The LLM then creates a plan with full knowledge of tooling, deterministic retrieval steps, uh citation filter, bounding boxes, and this whole thing. s69 00:06:15.920 --> 00:06:18.000 Now, let's actually make this specific. s70 00:06:18.000 --> 00:06:23.040 Let's say I'm saying, "What is the adjusted EBITDA of, you know, Google?" s71 00:06:23.040 --> 00:06:26.160 Think about what's involved in getting answering that question. s72 00:06:26.160 --> 00:06:31.600 I have to first retrieve a bunch of facts across 10-Ks, arguably over potentially different years. s73 00:06:31.600 --> 00:06:42.200 I have to figure out, uh you know, in an investor presentation and an earnings call, what new information was disclosed or if a restatement came out and now I have to actually update a number. s74 00:06:42.200 --> 00:06:45.240 All of that is what the model thinks through. s75 00:06:45.240 --> 00:06:54.240 From there, we have tools that actually retrieve, compute, build tables, and register this entire infrastructure into a citation registry. s76 00:06:54.240 --> 00:07:03.040 That citation registry is what allows us to not only give magnitude or metadata associated with numbers or facts, but type those facts. s77 00:07:03.040 --> 00:07:05.440 So, a simple example here. s78 00:07:05.440 --> 00:07:10.240 Let's say I do look at Palantir's 2025 10-K. s79 00:07:10.240 --> 00:07:16.880 In there, it says Palantir's 2023 revenue was, you know, let's say a billion dollars. s80 00:07:16.880 --> 00:07:20.400 All of those numbers that I just said need to be contextualized. s81 00:07:20.400 --> 00:07:28.600 That billion dollars, even though it appears in the 10-K from 2025, has to be annotated and treated as 2023's data point. s82 00:07:28.600 --> 00:07:29.840 I said billion. s83 00:07:29.840 --> 00:07:34.080 Well, if you looked at a 10-K or a 10-Q, it usually says one. s84 00:07:34.080 --> 00:07:37.640 The billions is up in the table somewhere far away floating around. s85 00:07:37.640 --> 00:07:47.880 So, we have to somehow take these complex objects and actually give them the annotations to be able to pursue meaningful and traceable output that's actually correct. s86 00:07:47.880 --> 00:07:50.720 Now, after the fact, models like saying stuff. s87 00:07:50.720 --> 00:07:52.280 That is what they do. s88 00:07:52.280 --> 00:07:59.120 So, what happens when a model actually invents a number that we cannot deterministically prove came from a trusted source? s89 00:07:59.120 --> 00:08:03.920 We have to have a post process a post processing step to actually strip this information out. s90 00:08:03.920 --> 00:08:14.240 And so, we do this and we give you the full traceability in the most I'm going to claim this and I think it's true in the most granular form of any tool on the market. s91 00:08:14.240 --> 00:08:23.080 Which means you can click on everything in the Kepler platform and I'll show you screenshots that actually allow us to prove this architecture works globally. s92 00:08:23.080 --> 00:08:31.840 So, um let's talk a little bit about how this works in the realm of Turbo Buffer who's uh been nice enough to give us a stage to have this conversation. s93 00:08:31.840 --> 00:08:33.479 First and foremost, let's look at a 10K. s94 00:08:33.479 --> 00:08:35.560 Let's say we see total net sales. s95 00:08:35.560 --> 00:08:40.000 That fact that $416,000 $416,000 It's a good point, right? s96 00:08:40.000 --> 00:08:41.280 We don't know what that means. s97 00:08:41.280 --> 00:08:46.800 That $416,000 number um needs to have facts associated with it. s98 00:08:46.800 --> 00:08:50.920 Things like a value, a unit, a scale, a period. s99 00:08:50.920 --> 00:08:52.480 Now, here's the problem. s100 00:08:52.480 --> 00:08:57.560 For those of you who work in this space, everyone can say XBRL solves all this issue. s101 00:08:57.560 --> 00:09:01.520 We don't need to worry about pulling information out of documents. s102 00:09:01.520 --> 00:09:06.320 That's true if XBRL weren't incorrect like some 10% of the time. s103 00:09:06.320 --> 00:09:12.360 But even more so, how many people are just looking at public information when they're making investment decisions? s104 00:09:12.360 --> 00:09:13.000 Fairly few. s105 00:09:13.000 --> 00:09:21.520 They have their own One Drive, their own BIP sync, their own various investment memos that they've put together and a slew of internal information. s106 00:09:21.520 --> 00:09:32.476 So, for our architecture to work, we have to work not just across 10Ks and 10Qs, but also on top of your internal data, your SharePoint, your Box, your Ignite, your One Drive, your Outlook. s107 00:09:32.476 --> 00:09:32.480 [snorts] s108 00:09:32.480 --> 00:09:37.720 So, we structure the data where the data structured and then we implement search where it isn't. s109 00:09:37.720 --> 00:09:48.120 And what that allows us to do is create a namespace in Turbopuffer that actually combines both the BM25 and hybrid search with the full guarantees of the Kepler platform. s110 00:09:48.640 --> 00:09:49.720 Here's how we do it. s111 00:09:49.720 --> 00:09:52.040 And this is probably the slide I know people are taking photos. s112 00:09:52.040 --> 00:09:53.480 This is the magical slide. s113 00:09:53.480 --> 00:09:58.320 I'm not worried about revealing this because this is a very hard problem and it moves quickly. s114 00:09:58.320 --> 00:10:05.240 So we use zero entropy so zero embeddings we use 1280 dimensions here. s115 00:10:05.240 --> 00:10:11.960 ANN is approximate nearest neighbors plus BM25 is how we actually grab information across I look at everyone's taking photos. s116 00:10:11.960 --> 00:10:13.087 This is amazing. s117 00:10:13.087 --> 00:10:13.720 [laughter] s118 00:10:13.720 --> 00:10:17.600 This is how we actually grab this information across our namespace. s119 00:10:17.600 --> 00:10:27.080 We do do reciprocal re-rankings to actually make sure that we pull the right usually 60K or 60 documents back and then we do another form of re-ranking. s120 00:10:27.080 --> 00:10:33.120 And this is a trusted form of re-ranking where we pull back the sources that we actually trust more than others. s121 00:10:33.120 --> 00:10:39.160 For example, different sources will have different pieces of information that claim to be the same thing. s122 00:10:39.160 --> 00:10:50.760 An earnings transcript an earnings call transcript where a presenter says my revenue was this that disagrees with exists in the 10K is actually a it's not a bug it's a feature. s123 00:10:50.760 --> 00:10:54.800 There's two different pieces of information I have to actually deconflict across these. s124 00:10:54.800 --> 00:10:56.720 So you need to be aware of that as well. s125 00:10:56.720 --> 00:11:02.600 We use a cross encoder to actually do the re-ranking and then we return that result to the model. s126 00:11:02.760 --> 00:11:06.800 This step in this process is predicated on two things. s127 00:11:06.800 --> 00:11:14.160 It's predicated on one our ability to rerun embeddings as new information comes in incredibly cheaply. s128 00:11:14.160 --> 00:11:16.360 So generating embeddings is not a cheap process. s129 00:11:16.360 --> 00:11:26.200 If you can do it and you can do it efficiently great but if you're hosting your own embeddings model it becomes unless you're hosting your own embeddings model it becomes really expensive and very difficult. s130 00:11:26.200 --> 00:11:30.160 So strategy here is we search wide then we re-rank. s131 00:11:30.160 --> 00:11:33.200 Now there's tenant isolation that also becomes really important. s132 00:11:33.200 --> 00:11:39.000 Now, when we actually hit any kind of a document, we have a bunch of different pieces that are important. s133 00:11:39.000 --> 00:11:44.560 For example, it would be very unfortunate if Citadel Desk 1 saw Citadel Desk 2's research, right? s134 00:11:44.560 --> 00:11:48.080 So, we have to have security built into the document level as well. s135 00:11:48.080 --> 00:11:57.600 So, re- The first thing we do is if you start at that middle query that goes to the document, um every hit is reauthorized in Postgres, meaning we check the citation, s136 00:11:57.600 --> 00:12:02.000 the value, the document across every one of these computations. s137 00:12:02.000 --> 00:12:04.080 Every builder adds a new filter. s138 00:12:04.080 --> 00:12:15.840 So, we're building up these complex filters over time that actually allow us to make sure that as we're querying information and querying data, we're not leaking data inadvertently across a bunch of different sources. s139 00:12:15.840 --> 00:12:20.960 Only Auth itself can create a scoped token to actually create this information. s140 00:12:20.960 --> 00:12:23.920 And of course, the user and aura come from a signed device. s141 00:12:23.920 --> 00:12:30.800 So, this is the way that we ensure security, and this is the way that we ensure your alpha stays within your firm's ecosystem. s142 00:12:30.800 --> 00:12:36.720 Alpha's an interesting question which we can talk about in a world where everyone has access to LLMs as well. s143 00:12:36.880 --> 00:12:37.680 Cool. s144 00:12:37.680 --> 00:12:42.000 Now, how do we make sure a number can appear in the Kepler platform? s145 00:12:42.000 --> 00:12:46.400 First and foremost, we assign unique identifiers to each fact in our platform. s146 00:12:46.400 --> 00:12:53.120 Meaning, we need to have identifiers that both can do entity recognition and entity resolution. s147 00:12:53.120 --> 00:13:01.120 Each one of these source IDs corresponds to a bunch of metadata in a fact stored in a in a deterministic engine. s148 00:13:01.120 --> 00:13:06.520 In order for us to show you the output in a way that's customizable, we need to have a display value. s149 00:13:06.520 --> 00:13:16.640 So, if you want your whole Kepler platform in thousands of dollars or in yen or in any of these, we need to be able to post process and show you the actual output. s150 00:13:16.640 --> 00:13:20.960 And we have to show you how we got it and where we got it from. s151 00:13:20.960 --> 00:13:28.480 So, the invariant in Kepler, you can click on every number, every fact in the platform and see the actual source that we got it from. s152 00:13:28.480 --> 00:13:33.920 And that's how we actually give you numbers that are truly citable, deterministic, and reproducible. s153 00:13:33.920 --> 00:13:37.160 This is also the item potent guarantee that we're able to give to people. s154 00:13:37.160 --> 00:13:44.920 By having this guarantee, we can run the same query over and over again wherever and get the same results. s155 00:13:45.880 --> 00:13:50.520 So, here is what we uh have to deal with. s156 00:13:50.520 --> 00:13:57.440 When we ran this whole corpus of data, um a third of all of our citations are the math that the model didn't do. s157 00:13:57.440 --> 00:14:10.000 What that means is pulling every one of these citations out, whether it's web results, spreadsheets, whether it's coming from a filing exhibit, the model itself can make up a bunch of things from its own corpus or its own s158 00:14:10.000 --> 00:14:16.360 um training set that are really, really difficult to work around, especially when you go to historical data. s159 00:14:16.360 --> 00:14:22.240 For example, looking up a revenue number from 2005 likely lives within the model's training corpus. s160 00:14:22.240 --> 00:14:29.760 But it if we retrieve it from that corpus, we can't guarantee that it was pulled correctly from the right year and right scope, and that we can actually trace it back. s161 00:14:29.760 --> 00:14:39.839 So, just off of a whole, a third of this are math that the model that a third of all citations are the math the models did not do. s162 00:14:39.920 --> 00:14:42.200 Now, how do we actually do this live? s163 00:14:42.200 --> 00:14:53.160 As tokens move forward as we're flowing kind of like we're doing inference, um we give these filters and these references to an LLM that has to actually resolve these. s164 00:14:53.160 --> 00:14:56.200 Think of this as the LLM dereferences the pointer. s165 00:14:56.200 --> 00:15:02.880 We then have a registry where the sources themselves are looked up, and we pull information out of that registry to give you the traceability. s166 00:15:02.880 --> 00:15:09.120 Now, the really powerful thing here is when we don't have a match, we strip that number. s167 00:15:09.120 --> 00:15:11.040 That's a very aggressive thing to do. s168 00:15:11.040 --> 00:15:17.760 It means your LLM will not be that wonderful thought partner that you thought it would be, where it would make up, speculate, and project. s169 00:15:17.760 --> 00:15:24.560 This means at this moment, and I'll say this, Kepler's not a super useful tool if you rely on model hallucinations. s170 00:15:24.560 --> 00:15:32.040 Meaning if you're trying to be creative or you're trying to I don't know, write marketing content, Kepler's not the right product to use. s171 00:15:32.040 --> 00:15:41.280 That means there are certain industries in finance where opportunistically selling, like banking, probably don't rely on numerical accuracy because your job is to sell an offering. s172 00:15:41.280 --> 00:15:46.560 It's not to create a DCF that's backed by accurate traceable information. s173 00:15:46.560 --> 00:15:57.440 So, by checking against our sources, we've effectively created a ground truth retrieval engine across any corpus of data, including public equities as well as your own internal data. s174 00:15:57.440 --> 00:16:00.160 Now, this is annoying. s175 00:16:00.160 --> 00:16:04.640 This means the platform cannot be as magically powerful as many people want it to be. s176 00:16:04.640 --> 00:16:09.560 So, how do we actually ensure whether we should strip a citation or whether we should try to repair it? s177 00:16:09.560 --> 00:16:17.200 And so, the core insight here is we have a post-processing step that actually tries to do a citation repair. s178 00:16:17.200 --> 00:16:28.800 Which effectively means if we find information that we can't reliably trace, should we try to repair that and do an on-the-fly agentic determination that says this information came from this source? s179 00:16:28.800 --> 00:16:31.400 The answer is in certain spaces we actually do do that. s180 00:16:31.400 --> 00:16:39.560 And that's probably the second piece of this presentation that we can keep a little bit, um, kind of like wishy-washy, which is how Kepler maintains its IP. s181 00:16:39.880 --> 00:16:40.520 Cool. s182 00:16:40.520 --> 00:16:45.880 Um, now, when we actually run these queries, there are some numbers that I think are worthwhile to share. s183 00:16:45.880 --> 00:16:57.320 Because most people in this industry right now are wrapping a lot of financial research with an LLM and they're saying, "Great, everything is right as long as things are, you know, ragged properly, I should be good to go." s184 00:16:57.320 --> 00:17:06.880 So, our filter across native, just like running it on Claude or ChatGPT, um, refused 10,981 citations from our evals. s185 00:17:06.880 --> 00:17:13.400 Meaning 10,000 about, I guess, 3,810 of these were just straight-up wrong numbers. s186 00:17:13.400 --> 00:17:16.720 This is the scariest thing ever if you live in financial analytics. s187 00:17:16.720 --> 00:17:26.160 This means hooking up Claude to uh, uh or open AI or any of these models to just a native rag tool or a vector database actually isn't enough. s188 00:17:26.160 --> 00:17:31.760 Now, there's really cool things that you can do if you're looking at just semantic search and that's it. s189 00:17:31.760 --> 00:17:38.160 But if you're looking at anything that requires numerical accuracy and auditability, you actually can't trust regular rag platforms. s190 00:17:38.160 --> 00:17:41.800 This is a scary thing for a lot of startups that are effectively chat GPT wrappers. s191 00:17:41.800 --> 00:17:43.440 I'm not saying that in a negative way. s192 00:17:43.440 --> 00:17:46.040 It's the reality of where we are right now. s193 00:17:46.040 --> 00:17:57.680 So, a new kind of architecture has to emerge and that architecture has to combine both the auditability and traceability with the deterministic elements of of course like numerical accuracy. s194 00:17:58.200 --> 00:17:59.880 Um models don't do math. s195 00:17:59.880 --> 00:18:03.880 This is pretty simple and I think it's actually pretty critical um insight here. s196 00:18:03.880 --> 00:18:15.120 Uh anytime that we see some kind of situation where the model is uh forced to do math or if it can't reason through something, we effectively make that impossible. s197 00:18:15.120 --> 00:18:20.320 By having that post post processing step, the model can't actually write something it can't cite. s198 00:18:20.320 --> 00:18:23.600 Meaning the model is structurally incapable and I'll be careful. s199 00:18:23.600 --> 00:18:28.360 The platform is structurally incapable of producing hallucinated numbers. s200 00:18:28.360 --> 00:18:32.280 The model can still hallucinate and it definitively does. s201 00:18:32.680 --> 00:18:33.000 Cool. s202 00:18:33.000 --> 00:18:40.920 Let's talk a little bit about what Kepler looks like and this is a tiny sliver of what our platform does when we show the deterministic traceability and all that. s203 00:18:40.920 --> 00:18:44.440 Um happy to chat through uh this as if it's if it's useful. s204 00:18:44.440 --> 00:18:57.360 But a simple question like Apple's 2025 revenue and how fast it grew, every number being clickable means by clicking on these, you see not only the 10-K or the 10-Q or the 8-K or the investor presentation that it came from, s205 00:18:57.360 --> 00:19:00.600 but you see the actual calculations that are done behind the scenes. s206 00:19:00.600 --> 00:19:09.800 Things like I did these two numbers and I have a 6.4% um increase all becomes present and traceable on the platform. s207 00:19:09.800 --> 00:19:13.480 So, this architecture I think will become the standard. s208 00:19:13.480 --> 00:19:26.360 Um, the opportunity in front of people that are building financial research tools or operating in the realm of AI is right now limited by the lack of architecture that actually supports the workflows you want to do. s209 00:19:26.360 --> 00:19:42.720 You can set up fully managed agents right now that try to ask for ideas based on, you know, generated information, but you still can't produce a DCF, which is why we still have analysts down the street from here or working right now on stripping information out of S-1s and taking information from a PDF s210 00:19:42.720 --> 00:19:44.480 and putting it into Excel. s211 00:19:44.480 --> 00:19:46.520 Our goal is to end that. s212 00:19:46.520 --> 00:19:54.640 And so, um, I do want to end on a kind of one very important thing, which is our guarantees do end somewhere. s213 00:19:54.640 --> 00:19:57.000 We make models site numbers, not write them. s214 00:19:57.000 --> 00:19:59.240 We verify every citation. s215 00:19:59.240 --> 00:20:03.800 Uh, we push arithmetic to the compute layer, not to the model layer. s216 00:20:03.800 --> 00:20:06.440 We isolate tenants and we measure recall. s217 00:20:06.440 --> 00:20:13.640 This means we can actually guarantee that the ground truth research that you do in your platform is unique and powerful. s218 00:20:13.640 --> 00:20:19.440 It does not mean that we can create a full-scale hedge fund that's replacing everyone, at least right now. s219 00:20:19.440 --> 00:20:28.520 And so, I see the move, and we'll talk about this in, I guess, 30 seconds as we go, uh, to the next place, but, um, the move that I'm seeing is folks will be more quantamental. s220 00:20:28.520 --> 00:20:29.320 This is the future. s221 00:20:29.320 --> 00:20:30.640 This is the opportunity. s222 00:20:30.640 --> 00:20:42.320 The more that we can use unique retrieval mechanisms in conjunction with, of course, deterministic retrieval, um, we'll end up in a really powerful place where AI can finally do the stuff that we don't want to do s223 00:20:42.320 --> 00:20:45.440 instead of sitting there and doing the ideating that we do want to do. s224 00:20:45.440 --> 00:20:48.600 With that, um, there's an appendix here, but thank you so much. s225 00:20:48.600 --> 00:20:51.064 Uh, I appreciate the time. s226 00:20:51.064 --> 00:20:53.064 [applause]