Collaborate on Digital Bill Format for Animal Science with Ease Using airSlate SignNow

Digital bill format for Animal science

[Music] hi everyone thank you for joining us my name is Kiera Pandya I'm a technical lead on partner engineering for machine learning at Google cloud and the topic we're gonna cover today is how do you automate invoice management right well you're every business practically has invoices that they're either paying or sending out and these are pieces of paper and how do you how do you automate that using machine learning alright so going back to the roots why are we here what is the goal the goal at the end of the day for Google here is to democratize AI we want to take all these bleeding edge innovations in artificial intelligence that our research teams Google brain are coming up with and making them available to your businesses right to to make it so that instead of needing one of the thousands of data scientists globally one of the millions of developers globally can then use the same technologies that's the goal here and with that in mind the the question we have to ask ourselves is what given the limited resources we have every organization at the end has limited resources how do we target our resources to get the most ROI for everyone involved not just us our customers our partners and what that leads us to is the question of unstructured data so one of the one of the very sort of interesting pieces of I guess statistics or whatever that a lot of people don't realize is that upwards of 90% of enterprise data is unstructured it's all those emails it's all those Word documents it's all those those text files it's all those PDFs sitting in accounting that's all unstructured data and the challenge with unstructured data is not only is it abundant it's hard they're you think a structured data you think of a database right you need average CTO goes oh yeah I have XYZ amount of petabytes in databases and the way you access that is easy you write a bunch of sequel queries or there's a standard ecosystem that everybody reaches for of tools but if I gave you a chunk of unstructured data and say go get me answers there's no clear answer there is no standard set of tools that you're gonna use someone's gonna reach for you no regular expressions or you know some of those things but what we're doing here now is applying em out and the idea is that machine learning can vary has it has a very clear path to making sense of unstructured massive volumes of data part two of the story is the transformations coming right businesses used to be run like this this is what we're moving to and you should have started your journey a while ago but if you haven't that's still okay you can still do it now right and the theme underlying team here is the announcements you saw around Auto ml right all these cutting-edge technologies you can apply them to your business to your existing business process as you are on this journey of transformation so the announcement this morning was about Iron Mountain and our partnership there and what are we doing here the goal is to advance automated document understanding we're working with Iron Mountain to with their customers permission digitized data and derive understanding from it so all that dark data that's sitting stored away from years of Records literally the US corporate history right of the Fortune 1000 extracting valuable today dark data using machine learning but as all these presentations go right the ml guy comes on stage and goes on about how ml is gonna change the world in the future we'll get there that's true but let's talk about here now how can you use the tools we talked about this morning and all day today and throughout next Auto ml and machine vision to to automate an existing process so this is about here and now this is something that you can start using today so as one of the scenarios in the thousands we could tackle with how businesses deal with documents the one that seems to impact the most number of businesses is hey I get I buy stuff as a part of doing business I people money or I sell stuff and people owe me money and these documents need to be processed we've been doing that since 1860 this is actually an actual invoice from 1860 and this is a modern one nothing's really changed and the reason it hasn't changed is the computer science behind making a protocol packet to have two computers talked about hey you owe me X is easy but there are too many players the structure of that packet cannot be agreed upon so what we've done is we've sort of fallen back to well we have accounts receiving an invoice the invoice team has people that look at it and approve it and it's a you know it's a human intervention process so let's change that right so what does it mean to do document understanding right at ten thousand feet the thing people usually jump to is oh this is OCR this is not just OCR OCR is optical character recognition literally the hello world of machine learning for any new learning data scientists is a dataset called M nest and this data set is understanding characters it's literally the hello world of the industry what we're talking about here is going from that extracted text to to recreating the structure that is a parent to a human that's reading the document recreating that structure and then putting it into a system that manages structured data and it's part of a business process right that is truly document understanding and at 10,000 feet as these three steps and where OCR is literally just step one okay so let's talk about each of the pieces step one you extract the text so google has the cloud vision API this API is solves computer vision challenges and problems that are general in nature right it's not going to identify my business's widget a versus widget B but if you want to you know identify a ship versus a boat right it can do that because there's a general problem it impacts everything this is the technology that backs Google's efforts in search part of this offering is an offering to do OCR so when you use Google search we're not just looking at the metadata around the search right we're looking into the images to find your relevant images so the technology that powers that is also what powers are the division API to do OCR so okay we have step one we give it an image it gives us the text part two of this is that the pricing here is a game changer I'll let the slide speak for itself honestly but it's pennies on the dollar and one of the keys sort of friction points in business right there's a reason it's hard to structure businesses on a one dollar transaction because the transaction fee completely kills any margin you have so as you're trying to reduce your cost of transactions reduce the friction of transactions this becomes very important so this basically makes the cost of transactions pennies on the dollar all right step two so we got the text out of the document now we got to restructure it we said we're gonna apply machine learning that's great now you're gonna need a data scientist and this person will go away for six months eight months and come back with a model in that period you will have spent lots and lots of money lots of lots of time lots of lots of engineering and lots and lots of organizational focus on solving this problem where each of these is a very complex problem in itself that engineers specialize in so we go back to our first principles my first principles are that's not okay we need to make this technology available and accessible to every developer so we have auto ml the idea behind auto ml is that all of this becomes one click it becomes one API call you give it the data you give it the data set and you tell it go train and in a couple hours it comes back to you with the model that will have best-in-class accuracy and then it gives you again a one API call way to say go host this in a scalable way for me so that I can just make requests to it and I don't have to worry about well I need the right infrastructure to do low latency inference on my model none of that you don't have to worry about it you just tell us and we host it for you so what does Auto ml vision do in this case auto ml vision is what's in the in the parlance called a classifier it builds you a classifier you give it classes of training data let's say hey here's a bunch of pictures of handbags shoes and hats and you train it and then at the end of the day it can it gives you a model that when you give it a picture of a shoe even when it hasn't seen before but it's a shoe a human would recognize that as a shoe it'll tell you this is a shoe in concept pretty straightforward in implementation billions of dollars of research so I said you know it's gonna give you one of the best most best-in-class models auto ml produces more accurate models with fewer parameters and practically all the other what I would call sort of benchmark vision models today so you've got inception before you've got ResNet 152 res next 101 BG 16 and you can see the red line with Auto ml on top where depending on how much time and energy spent on training it you get more and more accuracy with fewer and fewer parameters now why do that why does this matter this matters because because you I know what you're thinking but well if Google's gonna do it then it doesn't matter it matters in latency and compute cost at inference time so once the model is built you're gonna make hundreds of thousands of requests over time against it so if a model has ten percent more parameters in it you're doing 10 percent more compute across hundreds and thousands of requests so this is important the other reason it's important that's what that was sort of this axis the other reason it's important this axis is that if you get it wrong either your business process has a problem down the line or now you have to employ the person that you freed up to do a better task than just scanning an invoice is now back in the loop to make sure that the computer does screw up right so that's not okay - hence this is important so a little bit of sort of technical arada about auto ml so that the technology powering auto ml is called learn to learn where we've got a neural network that is effectively watching that the model built being built for you being trained and tweaking the training process as it goes so effectively sounds like a buzz word I get it but it's literally a model building another model is what's going on and that's called learning to learn and then along with that we're doing large-scale hyper parameter tuning to give you that that super accurate model without you having to do any work if you want to discuss the internals of this we can then probably here after - we can we can geek out so next up we got so we got the piece for extracting the text we got the piece for well vaguely we have this thing that does a classifier those are our two tools in our vision box how do we put this all together I promise that the solution here would plug into an existing business process right where you're not throwing everything away it's connecting to something real that you're using today so what I'd like to do is invite one of our partners up one of the ways that Google makes these things real as we work with our partners and our partners really understand how each business is special in its own way right Google Google builds the the surgeon's tool set if you will right to stretch an analogy and then our partners are the surgeons who take those tool sets and make them do the surgeries make it real and so in that in that vein I'd like to invite Chandra from Deloitte to come and talk to us about how they put these things together to solve the invoice management problem thank you yeah thank you Karen can you hear me okay okay good so our amazing technology right machine learning now with a new announcement of Auto ml from Safeway you must have heard yesterday so what do we why do we why does Google need partners right Deloitte as a system integrator we are very close to the business we understand each industry we understand the big different business challenges in every line of business and we take this technology and then we implement solutions which are very relevant for every every every company every industry every business process and the lot has been working with Google for multiple years now whether it's in the in the form of infrastructure as a service with us in the form of analytics using bigquery whether it's in the form of Adsense for creating a new marketing platform for our customers to get more insights and obviously auto ml before auto ml was announced it was available to us as a partner in a bit of beta form but we realized the power of ml just ml without Auto ml very earlier on and we realize that many of our customers can leverage a true benefit which is extremely cost effective it can be implemented in agile manner not wait for six months or a year to see the value so we started working with the googles machine learning team and the infrastructure team to develop this solution so let me walk you through it right so now every company right whether you you do services whether you manufacture goods you have to process invoices right to pay your vendors on time it's part of your and every vendor is unique because they all have their own branding they want they have a different structure different layout right so we need to be able to extract the invoice get the business data where it's a material number quantity amount address you want to make sure it goes to the right address so all this business information which is in different format for every vendor has to be extract in a very accurate form and before a financial document is created right so very crucial business process and let me share with you until now companies whether it's a midsize company or large sales companies spend millions of dollars with the commercial off-the-shelf solutions from different vendors to do the vendor inverse management which easily cost anywhere from two to four million dollar just from licensing now on top of that add few months of implementation testing with your ERP now every time a year B goes through a software change and you again have to go back to that inverse management soul do some testing do some remediation every time a new vendor is incorporated to your business you have to go through the programming all over again because they may have a completely different logo completely different format so it's a very very slow process very very expensive process and it does not leverage any benefits of cloud computing and machine learning and plus on top of all of this it has a very high sunk cost right whether you use all the features of your invoice management solution or not you have to buy all of it and you have to spend millions of dollars upfront so that is why we developed this solution Deloitte has developed this proprietary solution for automated invoice management for your s4 Anna ERP system so s4 HANA is sa piece large enterprise resource planning solution and we have developed this because we know ASAP s winona in detail we understand the business processes so it's going to help us obviously alternate your invoice management it's going to make it more efficient it's going to be much faster lifecycle from a solution perspective to ask an invoices post it to the business documents and obviously adopt it as you go so let's walk through a little bit in detail about the different components that are involved so as you see as you heard from Corrado obviously machine learning is a very crucial component along with the vision API those are the building blocks of scanning the invoice right to make sure we extract the right data so our developers who are very well form well-informed about machine learning and the coding required so we we develop our code around these two technologies then it has an cloud storage to store these invoices because every customer scans roughly you know anywhere from few thousand to several hundred thousand invoices on a monthly or quarterly basis a lot of money right and you do it most of the time in a batch mode right for most customers and the room for error is very small because you don't want to get an invoice inaccurately scan which which could be for thousands of dollars to a wrong vendor so there's a lot of money at stake it's a real business problem right so cloud storage helps us to store all those invoices for audit compliance perspect for much longer period at a very very lower cost and plus is available everywhere in the world as long as you have a cell phone or a laptop and obviously is for Hana which is your business system which will take that invoice right and then post it to the make - for financial document so now I'm going to invite Sudha or who's my colleague who's gonna decompose this architecture for you I'll help you through each of each and every process step by step how do we scan the invoice how do we get the OCR data how do how does the AP clerk gets to review that data and then post it in ASAP so that over to you thank you Chandra good afternoon everyone I'm so Dockery I am the global ACPs for Hana center of excellence lead at Deloitte and we've been working very closely with Google on building the solution so I wanted to take a step further get into the real details and kind of help you understand what the solution is I mean what components are involved details awarded and followed by a demo so it just cannot get started so there are essentially three steps three main steps to the solution the step number one is uploading of invoice it's kind of obvious right so you have your invoices which which need to be scanned and uploaded one important information for you to note here is that the invoices that are uploaded are stored in Google Cloud storage which is obviously running on Google cloud platform and the scanning of invoices can either be done from your laptops mobile phone or any device so all those capabilities are built in now the third important thing for you to kind of flow is when you when you talk about invoice management you're obviously talking volume so you'd every customer any carpet customer would have tens and thousands of invoices that get uploaded every day so so that's where the upload functionality comes in now next important part is the intelligent extraction of these invoices so you have got these invoices which uploaded and ready how do you actually understand the information there and there are again three steps to this intelligent text extraction the step number one is you should using the vision API to actually extract the detail from the invoice right so as for that we are using Google's machine learning vision API extremely capable API tool there we have seen it work with different kinds of invoices and when invoices often have scribble text text manually written all those kind of invoices also work extremely well with the vision API then the next component is how do you know now you have extracted the details how do you know what information is there so for that you vote you have we have yeah Mel config which kind of comes into play which helps as an understanding what information is where in the invoice so that's step number two now the third thing is you have extracted the information how do you know that this particular invoice belongs to a is a template from a particular vendor this how do you connect this that's where the third component which is auto AML comes in and then we develop the solution we've been developing that that over a period of time we had used traditional ml and we also used auto ml and once we started using auto ml we actually saw that there was a good increase in confidence percentage so when I show you the demo you would actually get a peek at it but with Auto ml we were really able to enhance the solution so with combination of these three extraction of the solution and then come classification of a solution you are actually able to take this unstructured data and convert it into a more structured data now the third component here is now we have extracted invoice we have all the information this information is posted into s AP and it's not just plain posting of invoice that I'm talking about they also business rules validations and checks that come into picture and close to kind of make sure that the data that is getting uploaded is in fact actually and is business-ready so those are the really the three components which come into play in this whole process so a morado kind of get you guys a little bit of a visual feel of of how this is happening so what you see in the background is how the text extraction is working this is where the visual API is looking at the components of the invoice and then it is extracting all the information which is there than invoice in different blocks as you can see over here okay so once the extraction is done then Yaman config kind of comes into picture this is like a really use really user-friendly and I would say easy to understand where you would you would look at the invoice and you're able to see what information is where in the invoice now let's get to the third part which is the cool part which is now you've got you have extracted the information on the invoice how are you actually going to identify that this information belongs to it is corresponds to this particular template and that's when auto ml comes in so as you can see here there are three different invoices tariffs completely different format from three different vendors so all this information is getting fed into auto ml and when a new invoice comes in you are able to identify that this invoice is corresponds to this particular template so I mean that's how the whole thing is working we have trained the model with with multiple invoices and really good results that we were able to see and that's what I wanted to show you once we get into the demo mode I think that's pretty much it let's so let's quickly switch to the demo yeah okay thank you okay so let me introduce the screen a little bit to those of you who are not very familiar with ACP actually a quick show of hands how many of you have used sa PRS be customers of you know how interacted with have worked with sa P okay yeah okay thank you but can I'm gives me an idea so so this is for those of you who do not know this is called s CP Fiore tiles these Styles essentially called us this is us a piece new user interface and the tiles are called as the fury tiles really convenient easy to use and we have developed this user interface using sa piece cloud platform so what you're seeing essentially here is s ap cloud platform there so the step number one that we have here is scanning of invoice like how I explained so we take these invoices we scan them this the demo would show you the invoices in online mode but this can very well be batched when we are using tens and thousands of invoices so these are the different components we have got the scanned invoice and I also wanted you to look at the Google Cloud Storage where this invoice actually gets stored okay so here is a look at the invoice that we actually gonna upload so I'm going to take one invoice we'll take it all the way through then I can show you another invoice which is in a completely different format as well so take the invoice a more clear or located so all the information there we have got the item details so we get into the the cloud platform screen a quick button there which we click on and we select the invoice and it's all successfully uploaded so once this is a uploaded you can actually go to the cloud platforms and you should be able to look at the invoice as well so this is what this is where it does all the invoices are going and sitting in the in the background next is a viewing of a scan invoices okay I want to pause here and kind of explain you this screen a little bit so what just happened is we had the invoice which we scanned there and put it in Google Cloud Storage now in the lookup in wire screen that invoice shows up right and what you're seeing over here is the extracted information from the invoice the third screen as a as I play is the actual image of the invoice so so once we click on the next button yeah so this is the header details of invoice there you see the confidence percentage 95 which quickly flashed there that actually shows that improved percentage that we got once we started using auto ml so this is the invoice that we scan all the information there and let me just pause over here so that we can actually absorb this information completely the left was invoice that God uploaded and towards the right is the actual invoice that that was uploaded and what you see here in the middle is the information that God extracted from this invoice and is now right over here right so this is the three-step very very I would say intuitive user interface another important point I want to really emphasize is I am showing you this whole entire thing in an online mode but when you are dealing with tens and thousands of invoices you would actually bash this so this is not I mean you you wouldn't really be clicking on a submit when processing these invoices what you will essentially do is we have the confidence percentages the thresholds which you can set for example you can say that anything over 93% or 95% whatever you may all those invoices would go through automatically if it's less than that then then potentially you can manually look at those invoices so those kind of thresholds and parameters can be sent to make a much more automated process so the POA information the details are kind of there that's the key yeah once we come back in the item information the confidence percentage again are really good there we click on submit so here is what happens as soon as you click on submit I want to pause here on a cut-off so this is the invoice number of the invoice which is actually posted in SA P right so this invoice immediately is kind of going in and posting in SA P so I said she'd what we do is the screen that you saw previously is the link now this is out of box F ap app Fiore app which is provided for supplier invoices so the inverse that we just created automatically got posted in SA P so getting to that screen and looking at this particular invoice let me continue so the invoice which we copied and pasted it there you see the document is already posted we can get into sa P you can look at the invoice the item details tax payment all everything got populated and even the attachment is also there so this is the invoice as you recall it is is what we had uploaded yep okay so I know that yeah so so the the previous screen that we saw was a si ps4 on a screen I know that many of you are ECC customers as well so we also wanted to kind of show this that the same thing the the invoice that is created is posted in the backend as well to kind of get get you to see how the ECC user interface or the backend GUI would look like the same invoice we put that in and you're able to get in into that and same information you see in the backend ASAP back end to the item details everything accurate you have the vendor details as well and then now we are able to look at the attachment so click on the attachment list it takes you to the attachment there open it and then there you go that's the image that you just scanned and uploaded yeah if we wanted to kind of get in and show the process for a different invoice as well I mean here is another invoice and I will try to kind of quickly run through so that you get a feel it off so this is the invoice again similar process getting in uploading the invoice and you would see that this this a completely different invoice with a completely different format again we were able to upload to the same application is able to read is able to understand right so this was pretty much it I mean this is this is really the crux of the solution if you want to see a live demo of it we are showing a live demo of it in our Deloitte booth you're most welcome to come in and then get a live look and feel of how the solution is looking so at this point of time I wanted to hand it over to kirat to conclude thank you sue that can we switch back to the slides please all right so what did you just see you saw in summary we did the three pieces right we extracted the text we restructured it back to some a form that is expected by again tying it all together to an existing business process system so the takeaway from all of this is that the technology you're seeing being announced all around you here it's all MLA I oh my god it's the future it's real today you can connect these like ml people like to do demos like hot dog not hot dog show you these abstract things right and my goal here if you take away one thing is that this tech can be applied to your existing business process today all it takes is a little creative thinking around okay well what is a painful point that I deal with every day that is either that either needs lots of human people doing a light touch sort of yes or no yes or no or is it just a very pro you know steps through process that just involves understanding unstructured data those kinds of problems are solvable today you just have to sort of sit down and think about wait a minute what if I were to apply ml to this problem how would I solve it [Music]