hi everyone this is rich and welcome to this webinar on how ml n a I can power small business applications in this webinar well start with a small introduction on Pitney Bowes and then talk about who these small businesses are what applications they are looking for and how algorithm has AI algorithms can be leveraged to build smart applications for these small businesses finally well have a Q&A session at the end of the webinar so for todays webinar our speakers are Ishida Akira.
Zhu whos a senior director for small business applications in Pitney Bowes name yeshua gangwar whos a distinguished engineer again at Pitney Bowes and finally John who is a developer advocate and a full-stack developer at algorithm iya over to you Shashi perfect hi guys welcome to this webinar again um what I did want to spend a few minutes on is give a flavor of who our customers are but before I do that I do want to talk quickly about Pitney Bowes for those of you who havent had a chance to look Pitney Bowes is a.
hundred-year-old global technology company we offer innovative products and solutions that enable commerce in the areas of customer information management location intelligence customer engagement and mailing and shipping solutions which is really our core we serve 90 percent of the fortune 500 businesses and more than a million small and medium businesses you know on the other end of the spectrum so both large and small businesses the small and medium businesses are really the crux of what we want to talk about today and let.
me start by just painting a profile of what these SMB customers for us look like like I mentioned we have a million plus or them so they do come in a fairly distributed vertical and segment but the few that I want to highlight today are in the verticals of legal services insurances health care professionals banking real estate and education just to give a better flavor of who these businesses are think of them as.
businesses with about half a million or more yearly revenue this is USD and about a business employee size of somewhere between 220 and 100 so its really not a really small mom-and-pop store but you know its an established business you know that’s now trying to be productive by using technology to automate some of their operations you know and functions within their office so what I did want to spend some time on is really to.
understand the psyche of our customers so you know these SMB customers – small and medium businesses what are their problems what are the problem statements that they have that we want to try and solve through apps as we try and build technology use cases using machine learning and artificial intelligence the libraries that we talk about later in.
this so lets look at the lets start with a small retailer so you know think of retailers who today sell on Amazon Etsy eBay most likely you know a small manufacturing unit or someone whos trying to put together a custom retail gifts novelties or even accessories I think for a whole lot of technology businesses so these retailers are essentially you know actively looking to grow and get new customers across there across their merchandise that they have and theyre also looking to automate away some of the functions.
like order management inventory management as well as think of how they do consolidation across all of the different workflows in their business if you think of healthcare professionals you know as your world you want you want them to be able to spend time on working with their patients and engaging you know and thinking about health of their patients their problem statements usually are in the frame of reference of invoicing and billing appointment management and schedule management and once they have those problems out of.
their way that apps can solve they can focus on really treating their patients and working with them the professional services business is you know think of insurance customers banking lawyers where again the primary use case for them to automate are around document management are around scheduling visitor management then all of the other good use cases that can help automate away part of their daily workflows more broadly you know SMB is today have been sold on the promise of technology solving for a whole lot of their you know manual.
repeted workflows and what we are really looking for our smarter solutions built through machine learning and AI to really help them unlock new revenue streams and focus on what theyre best at if you can go to the next slide so what I wanted to quickly run you guys through is as a quick summary video of the silca one for after a couple of minutes that showcases how we are looking to use apps to drive.
our business the Pitney Bowes cent pro c series all-in-one technology for mailing and shipping smart simple accurate we are building an ecosystem of intelligent applications for a c series device to assist small and medium business customers in their end-to-end Shipping mailing and addressing needs these applications will be available on our Pitney Bowes app store lets take a look at a few of these applications inform delivery for businesses the United States Postal Service is doing something new and digital with the mail for the first time small businesses to preview.
the exterior address side of letter-sized mail and track packages at one convenient location business owners can now track incoming packages and payments say for example if you’re an insurance agent your business requires you to deliver mails on specific dates ship miners will save time and money by using your calendar to create a daily shipping list for you it recommends the optimized shipping method to save your.
shipping costs click once to integrate it with your Google Calendar get recommendations and print shipping labels from calendar of events on the other hand if you’re a retailer you know how difficult it is to grapple with issues like returns and inability to update customers about their order status now with shipping companion you can take such worries in your stride this app helps retailers wholesalers and e-commerce firms at every stage of their.
shipping cycle by tracking all your shipments at one place in real time same-day delivery app enables simplified on-demand deliveries for SMBs on your sent pro CSeries device same-day deliveries are as simple as enter destination details and choose a suitable delivery option from a suite of multiple peer-to-peer shipping services the app also allows you to track your package to your final destination for small businesses the freshdesk companion app integrates your freshdesk customer support right into your CSeries device.
thereby allowing you to service customers faster with this app you can now view and update customer support tickets directly from the CSeries console improving support productivity the app automatically creates tickets for send post shipping alerts so you can proactively communicate with your customers as an e-commerce merchant you sell your products on websites and marketplaces or dOro lets you consolidate your orders from all.
platforms and manage them in a single view with just one click download your orders and print shipping labels directly from your CSeries device Pitney Bowes is bringing together an ecosystem of developers startups and enterprises to create digital first solutions for SMBs the Pitney Bowes platform perfected over decades is now open to bring the best of innovation to our SMB customers alright again that was.
just a flavor of how we are taking apps that loved to our customers through Pitney Bowes app store and if I had to use at this slide just to talk about a summary of what we talked about it is really that you know we have an open Android platform the device that you looked at in the video and what we want to really do is build an ecosystem of.
apps and connect the of 1 million SMB customers and all of their problems with app solutions built by you know a million plus after the first I do want to also take this time just to mention that we had a last Agathon late last year in q4 and we saw quite a few stat leveraged machine learning and artificial intelligence algorithms you know either through apis for data or through libraries with partners such as in Korea and Im quite looking forward to you know hearing how we can.
use these libraries to you know build apps that are faster more scalable I also look forward to seeing some of some amazing apps solving real-world business problems for SMB businesses coming out of this group that’s listen again and with that I do want to hand over to Joan from Inca Thea to talk more about their capabilities and how easy it is to use existing algorithms from algorithm eeeh around artificial intelligence and build your apps over to you John alrighty.
thanks very much so I am here to just show you around kind of what algorithm you can do and what we are so algorithmic is a marketplace with about 5,000 micro services in it some of those are developed in the house a lot of those are developed by academics or external researchers and brought into our system to cover everything from machine learning to various individual utilities and theyre really easy to use from any language one of the things we.
base our entire system on is that there’s a simple common syntax no matter what API you are using a new metal no matter what language you’re calling it from it always looks the same so its very easy to integrate into whatever you’re building we are also a platform for hosting server list functions so if you got something in Java JavaScript Python 2 or 3 our Ruby rust or Scala a lot of times you can simply copy your code right out of your current.
environment drop it into our web IDE or git repo and hit publish and you’re ready to go your algorithm is now running and available for other people to use you can even utilize GPUs in a serverless environment which no other provider does right now we have a 15 millisecond total overhead which is pretty small and a maximum runtime per function call of 55 minutes you can actually run some pretty long-running processes on our system and it costs about one thousandth of a cent.
per compute second and that’s bill to the caller the person is actually executing your code unlike something like lambda where you have to stand it up on your own billing account and then if you add another billing mechanism on top of it and of course because the syntax is common both on the client side and on the server side you can build on top of other peoples functions so if someone else out there has a cool class fire that does one thing you like you can wrap some more code around it and utilize their code inside yours right.
inside the server list platform seamlessly and then lastly we wont go into this now but for large enterprises that want either a private cloud deploy or an on-premise deploy we do support that where we take our entire system for running serverless functions and make it available inside their own private clouds that company can now automate the deployment and the discovery of machine learning models across their own.
enterprise but what I want to start by doing here is showing you what its like to actually use one of our micro services so Ill repeat this promo code again at the end but if you feel like following along you can go to algorithmic comm and use this promo code P b18 and what that’s going to do is drop an extra five hundred thousand credits into your account on top of the standard five thousand credits youll get automatically refilled every month for free so youll just head to.
algorithmic comm note that little promo code drop down at the bottom click that and that’s where you’re going to drop in your promo code pick any user name I like and any password you like and hit create your account well once youve done that youll see at the top there’s an AI marketplace link if you click on that you’re going to get a list of those 5,000 algorithms that are available for you to execute now theyre broken down.
by category at the top there’s also tags you can also search for them and that little search box in the upper left and one thing that’s unique about our system is all those individual algorithms are actually rated so you can see which ones are top-rated you can see which ones are called the most you can see which ones have been added and clicking into any of those individual algorithms its going to give you a description of what it does so there’s a brief description right at the top.
there’s a little Docs tab where it gives you a full description of all the inputs and the outputs of each function and then there’s a run example right at the bottom so when you find an algorithm that you think is going to be useful for you you can just drop whatever input you want into that input box on the left and hit that one example button its actually going to execute it right there for you you can see what the output is on your own data so this gives you a.
really fast way to go and test out these different algorithms without ever having even downloaded anything to your computer or installed anything now when you scroll down to the bottom of that page for any individual algorithm you’re also going to see cut and paste code in a variety of languages so you can actually simply run a curl directly from your console you can grab some JavaScript if you’re a web developer you can grab some Python and drop that into your script and youll notice the syntax is pretty similar across all of this I.
can actually demonstrate for you what that looks like so Ive got here one of our algorithms this is sentiment analysis and what sentiment analysis does is it simply takes some piece of text and it says hey is that generally speaking a positive or a negative statement so if I say hey I really like algorithm– yeah and execute that I get a pretty positive statement value over here but Im not restricted to just testing it out here I can actually go and grab this Python code down here.
there’s a nice little copy button to make my life easy I can step out to a console and one thing to take note of is at the top of each language installation box there’s a little install tab that just is a convenience for how to install our client library there now you can access everything we have through Rob Post if you want to but using our client.
libraries just makes it very easy to use this simple syntax so hip install algorithm yeah of course its already installed on my system so we can see the requirements already satisfied Im gonna start up a Python shell and simply paste in that code and we can see there’s an input document here saying whats the value Im going to be testing my client key for billing purposes the algorithm that Im executing and then Im going to pipe in that input and get back from our system.
right away a result hey that’s a pretty positive statement right there oh sorry there we go there’s the actual sentiment value so I can do that for any of my algorithms now to give you one use case example so analyzing sentiment of customer feedback which I just showed you positive or negative statements the input is well described here it takes a document and takes the text of a message.
notably it also has a list option so I could send in several documents all at the same time and get back an array of their sentiment values so I dont have to make lots of excessive API calls there’s our Python code sample that we just saw and then if Im going to use that in a production script a typical workflow I might use is maybe I want to.
know Im looking at my chat logs for the past history or Im looking at how people have reacted to my customer help and Ive got all that data either in a database or in a chat API such as say intercom I can grab each of those values out of the database or out of the API I can send each one of those to the.
sentiment analysis system or send them in batches and then I can do things like start correlated and hey does this operator tend to get negative or positive reviews when people are talking about its particular product do they tend to be happy or sad about that product I can even use it in real time to observe conversations that are currently taking place and find out.
whether or not I should be reacting differently to them so just as one very simple example we went and we grabbed an off-the-shelf chat bot tool we happen to use run Dexter comm but this could be integrated into anything and what run Dexter does is it allows me a simple way to create an interaction with a client now these are descriptive statements so if the client mentions some particular keyword I react with a certain piece of the script but what Im going to do is.
Im going to alter that a little Im going to actually inject some artificial intelligence into that automated conversation so the syntax for dexter is a little bit odd here but it should be possible to decipher what Im simply gonna do is when the user gives me some input Im going to send a post to algorithm yeah asking for that sentiment analysis API Im going to send.
it my authorization code and Im going to send it the body of the document and the body Im going to send is simply whatever input the user has given me in that chat interaction and Im going to get back from it a value which is how positive or negative that statement is so simply testing that out I can see okay if I say this is amazing I get back a value which is pretty positive if I say hey this is terrible I get back a value which is negative once Ive gone.
through and tested that then I start integrating it into an actual conversation so if I have a script which when they say hey what are your hours it says were open Monday to Friday 8:00 to 6:00 and they say that’s terrible my script can now notice that that’s a very negative statement and try to appease them try to say hey Im maybe not doing a good job of fixing the problem should.
I give you a call or something and then at that point I would transfer the chat over to a human operator because I realized that the user is getting irritated by the chats correct and its not satisfying their needs now this is one simple use case but of course I could go much much further with this I could add other algorithms for keyword extraction where it automatically figures out which topic theyre mentioning so even if I dont have an exact keyword match for what theyre.
asking about a I can figure out what the topic is that theyre interested in and still map them to the correct response I could have image recognition so if they send in an image of one of my products itll figure out what that product is and again direct them to the proper portion of the chat script and so on there’s thousands and thousands of natural language processing algorithms.
out there that you could inject into this type of chat now if you’re not a software developer if you’re not a coder or if you just want to try this out in a non code environment I also want to point you to a cool little add-on that somebody built which is our Google sheets plugin and you can find that a bitly slash Sheetz – AI and if you go there what youll find is a Google sheet that you can then make a copy of and once you do.
that it gives you direct access to many of our algorithms right from within a excel style context so now I can take a bunch of information that I have in some business environment Ive got to say a list of email addresses and I can say I want to inject an algorithm here and Ill pick there on the right Ill pick email validation and when I do that I drops in a little function then you can.
see at the top of the page here which just calls algorithm yes email validator and says whats right in the output box there hey yes that is a valid email address I can simply cut and paste that down on all of them and extract in real time what emails in that list are valid and which ones are not and then use that to call down my contact list and again I can do this with any data with any algorithmic algorithm that I want to run.
so now Im no longer limited to just writing code I can actually give this to someone in a business environment it was not a software developer and you can they can utilize machine learning directly out of it so heres a few of my favorite algorithms that you might be interested in playing with obviously there’s many more but just to give you.
an idea of things that I tend to use on a common basis parsing like parse face on the left is something that will break down the sentence for you into all of its parts of speech and so you can use this to say extract the nouns from a sentence or understand which verbs are adjectives there using it so that I can figure out what the actual meaning of.
the sentences and what I might want it to next sentiment analysis we just saw Auto tag will automatically extract keywords from text and figure out what the user is talking about and it does that with a broader understanding of the English language so its not simply a frequency but it actually understands what words are meaningful in the English language and which ones are not and.
which ones to ignore analyze tweets is similar to sense of analysis but it actually goes out and it scrapes Twitter for a bunch of information and gets you back sentiment but then it also gets you keyword extractions and other information or summarize URL which you can use to point to the webpage and get back a one paragraph summary about what that webpage represents in a human readable format so now I can generate Auto summarizations.
of say entire websites if you want to on the right Ive listed a few of the image algorithms I really like colorful image colorization is one of our favorites you can take a black and white image and it will automatically colorize it even though its never seen it before this works pretty well but the quality does vary depending on whether its seeing photos of a similar type before or not so try it out on your data set and see how it performs nudie detection I to V.
is an excellent algorithm for making sure that the content your users upload to a forum or to some other publicly facing image source is safe and is actually something you would want to show in public face recognition you can use to train and recognize individual human faces so you could do something like create a camera interaction where someone can walk into a room and itll automatically recognize whether or not theyre an employee of that organization and let them in or not or it could use it to look at your user base and see if.
you know them already a motion recognition allows you to look at a photo of a face and figure out what emotion that individual is feeling so you can do something like hook up say a camera on a cell phone app to see whether or not the users enjoying their experience with the application youve built an object detection Coco will go through an image and find in a in a data set of thousands and thousands of different objects actually identifying things in the frame so we can see in.
this photo on the bottom that there’s two people walking side by side that there’s a bus in the background and so on so stepping out of using microservices which weve just seen you also have the option to create a micro service and when you do that what you are doing is you’re putting code on to our servers which is going to be executed in that same way that youve just seen so I might have already written something in my favorite.
language whatever it is but I might need the user to execute it in a different environment for example Ive written a fairly complex R script which makes use of the statistical packages in R but I want the user to execute it from a way a char from a simple Python script and I dont want them to have to install R and all of the relevant packages onto their.
system well what I do is I just take my our algorithm I upload that code into algorithm use servers I declare what dependency packages it needs and then I publish it and then the user can use that service from anywhere from in a language at any time everything about that is going to execute on our server so its not going to create any CPU load on the end-users machine they can now do something fairly powerful even though theyre accessing it from a low powered device there’s GPUs available on the server and.
we allow many many parallel execution so you can have tens of thousands of people simultaneously using your code with no noticeable slowdown and when you create those algorithms you have your options you can either keep them completely private so maybe you’re using them only for yourself or only within your organization or you can choose to make them publicly available and if you.
choose to make them publicly available you have the option of simply allowing the execution to be public or if you like you can also make it open source so people can see your code and you’d be your option to either make it free or to charge a royalty on top of each execution and if you do charge a royalty every time someone else calls that algorithm those credits go back into.
your account and be a cache that when you reach a certain level now you can also of course combined with other algorithms so as a simple example if someone out there has built an algorithm which can identify what type of fruit a particular photo is and another person has written an algorithm which is can identify what type of vegetable a particular photo is I can then write about four lines of code to encapsulate.
the two of them and now someone can give me a photo and I can tell them its either a fruit or a vegetable and which one it is by simply calling out to those other algorithms and combining their output so I can do that I can also of course encapsulate my own very complex or secret logic that I wouldnt want to actually ship out to a client machine because then I can have that completely private on algorithm use servers so people can only execute it they can.
never actually see the code the syntax is going to be exactly the same whether you’re calling from a local environment or whether you’re calling someone elses algorithm from within your own micro servers that youve just created just to give you a very quick walkthrough of what that looks like when you hop onto algorithmic comm again instead of going to the AI marketplace now what you do is you click that plus in the upper right corner that’s going to give you an option to create a new.
algorithm you can give it any name you like and youll notice that its prefixed by your user account so in my case Im JPAC so I see algo : J Peck and then the name of the algorithm are in creating pick which language you want to write it in whether or not it needs internet access whether or not it needs GPUs etc and when you’re ready you create your algorithm now once youve done that you have the option of either.
get cloning it because every new algorithm you create gets an automatic get repo in our servers so you can clone that down and then use your own local IDE to edit that code or you can choose to use our web IDE which just brings up an interface where you can simply cut and paste and itll do syntax checking and the like right in real time on the.
web for you well once youve put in your code simply hit the compile button in the upper right that makes it run in our servers and then once that’s done you immediately can test it out on the console at the bottom so in this particular case Ive just written a very simple hello world algorithm Ill put in some Ill put in my name itll echo back to me hello John.
obviously that would include much more complex code in most cases lastly I hit the publish button in the upper right and I have all those options so I want to make it public or do I want to make it private do I want to put a royalty charge on top of this or make it free for everyone to use and note that you.
will also get to set a new version number so every time I publish a new piece of code it gets a new version number and that means that I can guarantee that my code isnt going to break peoples external dependencies in the future because they can always go back and reference an earlier version of my code if they need to and that gives me the flexibility to change it in the future however I want without worrying about damaging running apps once Im done I hit publish and that makes this then available for other.
people to execute if I have chosen to make it public I get the same input and output box someone hits run and now the code that I have just written is executing on algorithm– es servers again they get this down at the bottom different language options they can pick corel JavaScript Python etc copy and paste that hook and boom someone else is running my code right away so this is a quick wrap up if you want to see more demos head over to demos algorithmic.
calm I got a bunch of really cool stuff there including some fairly complex stuff like how to colorize entire videos for example or recognize objects in that video stream also some time serious information which will be really cool dont forget to use that promo code algorithm EUCOM if you do try us out so that’s P b18 and that’s gonna dump an extra $50 of credits into your account and lastly if youve got any other questions beyond this webinar feel free to hit me up anytime on JP e ck at algorithmic calm or if you go to the.
algorithm eucom website in the lower right corner youll see there’s a little intercom chat feel free to click on that and just ask for me and with that Ill hand it back to my colleagues thank you yeah so I think thanks John for showing how easy it is to use AI and machine learning algorithms using the algorithm ear marketplace let me give me give you a small demo on how these grid algorithms can be used to create an.
application for small business so in todays world social media provides a powerful tool for outreach and even some small business have started utilizing the power the power of social media in the age of Yelp and Google online reviews and the feedback has become more important to capture new customers so like take an example right these days we go to any restaurant or any new place we take a look at the reviews right so social media reviews have become really really important and then managing the online presence for SMBs have always always been struggle so using the power.
of Twitter apis and power of Tamiyas sentiment analysis api you created a cool demo to showcase so basically I created a small Android app which basically scans for Twitters foreign handle called my cool store so I just click on scan social media so this to showcase what it is is Ive created an example store handle and it has a bunch of boots.
so if I switch back you will see that I got the post I can see it to positive post to negative cause for asking one negative post and one neutral post so all this scoring is coming from Tamias systems and we are leveraging this name to show the insights to to an SMB user right so if lets say I am in store owner and I am worried about my online reputation and I want to manage that so this gives you insights on what people are writing about my store and helps manage them.
better so that’s I think what we have used in this app is again some Twitter apis we are using Peters SDK to fetch the results from all these API is always showing all the – its for this handle and then we are loading them and pushing them for sentiment analysis – algorithm yeah so literally what we are doing is that fetching all the tweaks for this store and then running it through.
algorithm he has sentiment analysis and creating a score so just see see the simplicity on how easily we are able to create an education which gives power to manage the social media presence and then atom store owner rights as an SMP customer I can see what people are saying about the business brand on the social media channel do market research and how people feel about competitor so I didnt.
demoed that within winded but literally if I know my my competitors I can get how their reputation is and therefore are talking czar and analyze the impact of marketing campaigns so lets say if I tie up with any marketing campaign and I want to see the impact that is having on the on my digital or social media presence that’s what it does so as I mentioned what Ive been using is fetching the tweets for the store running individually to Mias API is and just using that data to showcase these so we use to me as a rest-based API is.
for Java implementation in the Android and we are just using the rest implementation here the the number which we receive the number the sentiment score which we get from algorithm here so I think this for demo purposes I am multiplying it by 50 and we are basically using it to identify what is a bad score for me versus what is a good.
score for me and then I think to showcase the really how we can do an interaction that lets say if somebody has done given a bad review right so like for example this review is bad I can do some progress some action items are reaching out to the customer responding them and manage what kind of response we would like to give so then I.
think adding these additional dimensional dimension to this is that using Twitter we get the location address as well so might not latitude and longitude corresponding to that to it and using that we can define a pattern of places where the bad reviews are coming from so and this will go for a for an SMB which has multiple locations and you can see that model places the bad reviews are coming from and then I think you can manage and.
reply on the tweets or itself so basically this app can be used to search social media posts so I think the same thing can be extended I gave an example of Twitter the same thing can be extended to search on Google reviews Facebook and I think we can literally scan through these data sets and run the sentiments over there we showed the calculated calculation of.
the sentimental score detect social media incident reporting and generate alerts we can summarize the sentimental matrix and then viewing the post of the sentiment so I think that’s basically it so you can see how easily the API is of helped us create a very simple but yet a powerful application just by merging to the API is to create an online presence application so I think on the same line we have we have a bunch of other use cases which I think as an hacker as I.
think John mentioned and I think says he also mentioned it briefly about the upcoming hackathon so these are some target use cases which I think you guys can focus on a few are planning to join the hackathon so AI for competitive analysis I think I gave an example of if you want to see how my store social score versus how somebody elses my competitions social score and we in the same way I can be used to do more.
business insights then I think bought powered customer service solutions are becoming really really key and I think Jon briefly touched upon how easily you the same sentiment analysis we can create an powerful tool there and I think we just touched upon one algorithm there I am sure that we will have more API is in the marketplace which can help generate better BOTS and as you know I think all the stores and in fact all the businesses are looking for enabling a chat bot for I think whatsapp there.
having bought powered sort services for Facebook and I think anywhere and everywhere right so then I think AI based marketing and sales and I will enablement so I think this has been key challenge if you see this SMB which says he talked about they do not have such big media spends and they do not run bigger camp and big campaign so I think if AI can help understand what customer segments the the the store or the SMB.
should target then I think this will really be a big return for SMB so that’s a big claim and I think this is something which we are also looking forward to see as some hacks then intelligent CRMs predictive cash flows play and predictive not only creative with cash flows we can have predictive store sales or store wizard so take an example that there is an here saloon right so based on the previous data or.
previous users information we can create a very predictive model that during the day of time how many customers are going to visit how many customers going to visit in during the weekend the holidays so that as in small store or I can have the staff available to support that kind of up for uptake which might happen during holidays and other seasons then I.
think also business productivity solutions so I think SMB is again they are small start they do not have the strength of an enterprise having multiple workforces so they do multitasking right so I think doing lot of things like somebody sitting at the reception might be doing visitor management also managing the time entries so the productive applications for these these target audiences for SMB will also be something which well be looking for then I think.
again I think well call all the developers for our next SMB challenge I think we are opening the registrations for this on 6th of April again this is a global SMB challenge the themes for which will be SMB productive apps ecommerce enablement post and please shipping services international presence and open innovation for SMB so I think as you can see at the bottom we have bunch of partners and I think all the algorithm we are presidented that there.
you can use their promo code to get access to their API is I think almost double what they are offering so I think that brings us to the end of our presentation and well be taking questions and answers so I believe I think questioner answers we will take you can type it in in the quotient section and I think well be open to answer that son thanks everyone so if you if any of you have any questions please feel free to type in your questions and we are here to answer your questions and if not.
you can also email us or you know drop us a message at developer at fiba com and well be happy to answer all your questions all right so weve got a question about the future of AI in terms of financial tech and well I dont know if I can tackle that in complete detail I can definitely mention a few things there one is that when people think of AI in terms of financial tech a lot of times people think in terms of stock trading and the thing about that is that.
most stock trading occurs extremely close to the actual trades so you’re talking in some cases nanosecond response times so you really need to be on device and very local and a lot of time that requires a really elite data set so I dont think there’s a lot of opportunity as people seem to be under the impression that you can drop in some AI from those downstream data streams that you’re seeing that are seconds or minutes away from the actual stock trades and make any improvement or edge there that in practice doesnt tend to.
work what you can do is you can start correlating external data sources with financial tech in an interesting way you can look at some of the obvious stuff like whats happening in terms of coordinate early reports and doing natural language analysis on them and starting to understand ahead of time that a company may be improving or tanking based on what you’re seeing Ive.
heard of people doing things like taking call transcripts and running them in real time through AI for word recognition for understanding and trying to predict from that where their stock might go and of course there’s general correlation between companies things like looking at s1 reports looking at real time interactions with your own products and what customers are looking at or asking about in terms of other products and how those might relate is weve also got a request to look at some real-world examples built using.
algorithm– yeah and I think I can take a minute there to actually just show you some demos which might be even more instructive because I can I dont have to worry about any IP restrictions or anything of the sort so if you could toss presentation over to me please I will show off our demo gallery here alrighty so this is if you go to demos dot algorithmic comm that’s the easiest.
shortcut to get here and you can see some of the stuff weve done in real time Im certainly there’s whats very popular are image colorizer a significant number of hits on our site or just people using this for the fun of it and this is the thing I was demonstrating earlier where its seen a bunch of images already it hasnt been seen these specific images but it has.
seen similar images and so it attempts to figure out automatic colorization of those images and a lot of people just use this as a service there’s a couple of iOS apps out there where people can upload their images get them colorized back and we get a lot of people really saying hey this is awesome because I know Ive got this photo from the 50s of my parents that Id never seen in color.
before and it really feels realistic which is great but there’s some other really cool stuff here so if I look at say time series analysis that is a pretty nifty one and what we did here was we went out and we got actual use permit applications and filings from various City data on socata open government so this is basically going out to individual cities finding out what construction permits have been filed at different times of the year in.
those cities I can pick it pick a particular city I want to look at and then heres where the AI comes in I can start running for instance outlier detection on that I can run it through say a seasonality filter or I can go and pull out a forecast smoothed out on that seasonal filter so now I can really see over time what do I expect the permitting in any given city to be and I can use that to predict.
in this particular case it might be a good time to build when might not be but of course you can use the same sort of thing on any data and you can see that its actually doing all of this in real time its actually taking that data and transforming it against our API is but then pushing it back into this front-end so people can visualize what those.
results are going to be and then lets able to show you a couple of cool video things so for example there is the card detection algorithm this is one of our more niche ones so this will take you to tell you the make and model of a car and the algorithm itself is set up to take an image and then give you an answer what kind of car is this but what we did is we hooked up a video stream to it and.
because algorithmic allows us to parallel eyes all of those calls what I can now do is I can send into that algorithm all of the frames of this video simultaneously so I take in the video I split it into individual frames individual images and send all of those images to the same algorithm but that processing runs in parallel so its very.
very fast and out of that I can say ok at this point in time when I a granade the results back together heres a Hyundai hatchback and I can see the actual output from the algorithm right here one thing to make note of is that with machine learning any given data sets only as good as the data its been trained on so I can see down here it.
says if there’s a citron van but if I click on that what I find is that actually at Ram now this particular machine learning model if you go and you read about it is a car make and model identifier it knows nothing about trams so it certainly makes its best guess but it doesnt get it right fortunately you can actually see that in.
the output so it says the confidence is 0 08 which is very low one would be perfect confidence so this classifier is great for doing cars not so good for anything else lastly I want to show you the video transformer so again I said I can take a video I can split it into frames and then I can run an image out with them on each rain so here Im gonna run colorization.
on an entire video so for every single frame that’s running it through the colorizer and assembling this video in real time for me to see the results up now Ive got some flicker there I want to do some sort of interpolation between frames to smooth out the resultant video but it does an okay job it gets the skin colors right its got that classic keyboard beige but I can do other things too these videos too I can.
run through say a saline sea detector which identifies the point name which which would be of most interest to a human looking at them which is usually to say animals things moving peoples faces etc and I get this sort of heat map output of where each thing is that’s moving around so I think that’s where Ill cut it with the basic demo stuff.
and let me check back in to see if weve got more questions here mmm so we do have a question about running algorithms on the edge device we do not do that right now everything that we run actually occurs on algorithm use servers so if you’re for example offline then that’s not going to work out too well for you but if you are online even if you’re on a relatively slow connection we do have that 15 millisecond and see.
so we find in practice that even for very fast real-time applications were actually able to execute quickly enough that people do not notice and then lastly there’s a question about retraining so Im guessing this comes from a data scientist and what we specialize in this algorithm hosting we dont specialize in training that said there are two of our algorithms that do allow training which I could really show.
you very fast and what these are are that weve built these algorithms so that they have two modes they have a training mode and they have a prediction mode so unlike many of our algorithms when you actually send in a request you declare in that request hey am i executing this as a training step or am i executing it as a prediction step so I send in mode train along with my training data I do.
that many times for my new inputs and take note that I also assign it a specific namespace and then later on I can make a prediction against that in same namespace and get a result and we basically have two algorithms that do this one is this document classifier which allows me to take a bunch of documents and train it for specific results so this might be again customer.
help chat customers send in a bunch of questions and then in my chat tool I have tagged those questions with this was a question about product day this was a question about product B etc I would send each of those Val each of those chat contents and the tags that Ive assigned them in is my training data and then later on when someone came along with a new chat request and they gave me some paragraph of text I could.
send that in in prediction mode and it would give me back its best prediction of what keywords that what products that related to so that’s the document predictor and the other one we have that I wont show here is the face classifier same kind of thing I send it a bunch of images of faces with names associated with them in training mode and then later on I say switch the prediction mode I send their new face and it gives.
me back its best prediction for what kind of face that is were getting pretty close to the end here so I think Im gonna switch back to our presentation let me just do one let quit last check for questions okay and I think that’s what Im gonna wrap it up there so maybe my colleagues will want to do the closeout at this point yeah is there a slide on the just talking about.
the hackathon I just wanna highlight that the one part that I should have talked about is a hackathon signups will start this Friday the hackathon itself is global so youll have the opportunity to work alongside you know hackers and Im the person across the world I didnt the one part that I didnt get to cover around this was the apps that get submitted through the hackathon can be monetized as well so we do have an incubation program that allows winners from the hackathon to work closely with us and build out fully Production apps.
that then go to our customers we help with early market access as well as various models to be able to monetize that app both from an IP point of view as well as actually running it as a business so I do look forward to a whole lot of you registering i building apps with machine learning and artificial intelligence you know and i you know i look forward to seeing apps and winners from the set so Ill Ill stop there all the very best john this was this was.
creep III havent quoted in a while but I went back and signed up for all girls Mia and Ive decided to build my first app and hope everyone who listens in is also doing that right now thank you so much we will see you all very soon.
Leave a Reply