Legal AI + Finance = Noxtua

[Music] hey everyone I'm Le I'm the founder of Kain and very happy to present noxro to you one of the fastest growing legal language models in the world um we are a spin-off from Ox University we were founded in 2017 and mostly try to cross the borders between privacy and and AI especially when it comes to legaly we've been featured a lot uh in the everywhere in the news like uh especially in Germany Europe uh like Financial Times uh efforts at hles blood Etc um and what is Sayan about and especially nox we especially try to um see or address how we can solve um one of the main problems when it comes to AI in law first of all traditional language models like GPT have a lot of shortcomings when it comes to the actual quality of the model especially in very dedicated jurisdictions like for example Germany France Italy Japan um when it comes to really answering truthful um answers but also understanding how the legal systems work there because most of the data in these jurisdictions is not public so it's very hard to train models in these areas and you can only do so when you have enough Partnerships the second point is that these jurisdictions often have um very very strict compliance Frameworks um for example in Europe you have in many countries uh uh the professional secrecy um uh amendments that you have to sign which means that for example you can't process data uh on us clouds or sending data somewhere else where the data is at risk to be sent outside of Europe but more so also as a um provider of the service you could you you can never actually anywhere see the data that you process so everything has to be blind so to say and that's of course quite tricky um because you need a lot of security around it you have to prove also that um you never see this data and in Germany for example um and same also applies to other countries like France Italy um this is actually regulated under criminal code so if you actually ever saw the data of your customers this is not just something with a financial fine or so but this actually means that um this is at least couple of years in prison for the entire sea level so how did we solve this problem um well we actually trained a dedicated legal language model um for Europe and also of course extendable uh to other jurisdictions in the world um we cooperate with multiple different companies here for example CMS which is the largest law firm in Europe with around 7,000 lawyers but also other legal databases Publishers and so on so we together build the legal AI Alliance which is kind of like a data room for law um where all of this kind of proprietary data um is handled and is um collected just in order to train nox as a legal language model so nox is at the end and in the core developed with lawyers but especially also for lawyers and the second thing is that um we of course have to apply to as mentioned previously to a lot of um regulation here so um we have to actually make sure none of the data is actually ever leaving uh the country and therefore we also went into Partnerships with um companies like deuta Telecom for example um and we run the entire service um in conf or with confidential compute servers which means neither do Telecom nor we can actually see what is processed went through a lot of different certifications around five ISO certification C5 certification and so on um so what is nox Tu in essence about um it's mostly three factors we do legal research so coming up with very precise answers for um legal problems um including subsumption we have the ability to drop Clauses draft contracts but also analyze contracts um which is one of the most powerful features and I'm also going to show in a second um so we deliver a per hat assistant for lawyers and legal departments but also um we allow for others to use nox dra either via our API or connectors and for that we for example build a power automate connector or brighter connector so you can actually as a legal department um automate entire legal processes in the background um check Mass contracts check basically your entire company um for um infringements for example for risks Etc um and also we have um partnered already with um a lot of different uh legal Tech providers that actually integrate nox in their products like for example claw spas but what is about um the actual performance of the model and here we also see that um uh this is uh an example of two of our models uh nox dra Voyage which is um a legal search model and nox dra llm um how this specific training on legal data actually allows for outperforming uh the top models from open AI in this case embedding V3 and gp4 for these dedicated tasks so smaller more dedicated models actually can still and even more so nowadays outperform the best world models because you actually never see this very prioritary data in these models and the same holds of course for very uh for all sorts of other Industries like medicine and so on so we actually focus a lot on this ultra high value data set for training nox and this data is really of the ass here so the data set overall is of course much smaller than the data set that for example gp4 is trained on however we have a lot more legal data that these models are trained on and because we're not so much like influenced by all of the other things that are included in these models we actually achieve a much better performance so how do we train it we actually train with reinforcement learning so we have a dedic dedicated training environment together with Partners where they see um different outputs of noxr and including also their own outputs so it's a mix um between generated outputs from the model and human written output but the actual users the Train the training users I mean don't ever see which is human and which is um generated by the model they rate it and thereby we basic basically go through a lot of different sample reviews and over time this data has been grown to like right now more than 200 billion um token data set so um I'll also quickly show you an example um of nox tour um so of course you have all sorts of ability similar to chat gbt but the more interesting feature is that you can actually go through Mass reviews of contracts so let's say you want to go through DPA check then you can do so in a matter of a click so basically what I did is just using a Microsoft DPA the model knows exactly what to do here and it basically goes through an entire DPA in a matter of seconds and then you can of course download this as a report and what a lot of companies do is that they actually uh address this as a background process so it runs all the time and you can for example send all sorts of different contracts there um and then they of course get stored somewhere and uh for example SharePoint or whatever and uh then you actually uh automate a lot of the tasks that you have done previously by hand yeah I hope uh this shows uh some of the benefits uh of nox draw which I think is pretty impressive and is already helping a lot of different companies especially in the financial sector including companies like aliance Banks stock exchanges and so on which of course sit on thousands of different uh legal tasks every day thanks a lot oh [Music]