Streamline Sales Forecast Automation for Technology Industry

How to create outlook signature

well thank you everyone for joining I think we should begin here we did have some people just log on in the last minute or two welcome my name is Kjetil Enza today's webinar is AI machine learning for sales forecasting we're excited to offer this webinar this is the first in a series of webinars that we will be doing on AI learning that is excuse me AI machine learning San next slide today's speakers today's presenters are Sam Kahn he's been a practice lead here at Alpha bald for last year he is a practice lead for AI machine learning and Internet of Things he has 15 years worth of experience in cloud and artificial intelligence it's very very knowledgeable are obviously resident expert here at Alpha bald so this should be a really interesting webinar with him presenting because he pretty much knows everything about it my name is Jeff lens ax I am a client Engagement Manager here at alpha bolt I have over 10 years of software sales experience and in addition we have ty about week unfortunately no picture of daya because he just kind of hopped on here at the last moment he is the VP of consulting here at Alpha bolt and he will give a brief overview of sales forecasting so everybody has kind of a better understanding as to what sales for forecast and consists of next slides him today's agenda we will be covering the AI journey what that looks like obviously sales forecasting some of the challenges and insights of sales forecasting we will dive into time series forecasting and go over some of the popular techniques that people and excuse me businesses use when it comes to time series forecasting and then we will move into machine learning and kind of review some experiences our excuse me experiments and then finally talk about Microsoft's as your auto machine learning and give an example of that and then finally we will move into a question and answer session at this point I'd like to do introduce taya Bali who is going to review sales forecasting with us kaya hi everyone this is a yeah I hope you're all doing well and staying safe so just all of that I wanted to I'm pretty sure all of you know what I'm going to say but it's more about introducing sales forecasting from the perspective of the floor of this webinar so very simply put it is just a minute to predict future sales and their multiple ways to go about that and we'll talk about that in a lot of details from the course of this webinar but for this light most important two types of sales forecasting it's going to be short term and long term so a short term is mostly used by organizations to handle their productions as it is in place and long term is primarily used for strategic planning and being able to see how they want to plan on organization's growth so as you can imagine because sales forecasting is going is playing a role with future forecasting and setting organization goals it is a very important aspect of overall strategy for an organization and hence it's very important to have some sort of excuse she tied to your future sales predictions it's important because mostly your revenue targets are defined by your sales forecasting and then it is the driver for your costs on your workforce your cash flow management all of your resource management so it's important to have some accuracy tied to it and have some sort of phase or crusting the pour test that you have for your organization so with that I'm going to go on to the next slide and talk about the fact that traditionally speaking the basis for casting is done is that you look at the historical and you try to predict the future and that poses some challenges which obviously have to be mitigated if self forecasting is going to play a pivotal role in your strategy or your organization so the challenge is that it poses is for example let's say if we look at this historical and we looked at the sales of a product in Christmas timeframe what Thanksgiving thankful then there would be a spike there so as you would imagine we cannot use that November and December sales to be able to predict January February and March so traditional just looking at historical sales forecasting would fail in both instances similarly let's say being on a product and there is a huge discount being given because we have said launching it in the month of June and it's the end of the fiscal year and our sales team has to meet their sales quotas and this cannot be used to predict sales for the next year as well as well as the rest of the months because if I use June's numbers to predict July August and September it won't be realistic and accurate extrapolation of that data some more examples if you have two similar products that are launching two different ones one is go insane choose the other is low exchange December and the seasonality has an impact on the sales then again we have to take those things into consideration as we try to do the forecast of the sales forecast for that product so these are mostly internal factors but then also some external factors that are equally important to consider for your acrylic sales forecasting and from that front for example if there is a competitor product that is standing in the same time span this can be one of the reason for your product to have lower sales during that same time so you have to consider that as you evaluate the future sales of your product and then obviously more holistic event such as the session and disasters medical emergencies like the one we are going through have an impact as well so how we go about mitigating these risks is where di tells us and for that I am going to hand it over to Sam to talk and walk us through some of those examples and show us some demonstrations Thank You Saul most I have for setting the stage here so before we get into some of the details regarding AI it's important that I take a moment here and talk a little bit about what AI can or cannot do first off air is not a silver bullet or the auto magic cure and by that what I mean is that it's not a solution in and of itself it's actually a piece in the puzzle and in most cases as a strategic puzzle arm that can actually give a competitive advantage to be implementing organization so since it can provide the strategic edge therefore it's imperative for your organization or company to set clear goals on what it wants you to do to begin with so that actually happens when the due diligence is performed internally some of those things that I was mentioning earlier x2 and and also revisiting how the business is currently been conducted and you know what should be the end state and we're gonna we where we want to go to essentially so setting the goals is super important right from the get-go so once the goals are identified then it's a turn of investing into gathering the representative data so long established companies have data available but you know they have that advantage but that data is not unified and scattered across and different islands if you will I mean for example some of the data is going to be a developer in the finance application some sales or applications that operation and stuff like that but it has to be unified and the gaps in the data have to be filled so that you know you can come up with the representative data set that can be fed as input to be AI modeling process so so for a six so the lifeblood of a successful AI model is the quality of the data on which it's strained and it and it keeps learning by the way over a period of time and you know it starts providing reliable results and predictions only if the data that is P by 2 it is a representative start all right so moving on it's a turn-off identifying the technology stack and the solutes for the solution implementation you can bring in the experts who could help you and guide you through that process and lastly this is the most important point um you know you have the goal set the data is there that the tools and technologies have been finalized so rather than going all in for experimentation it's almost always advisable to create those goals down into smaller goals and objectives so you can pick one of those and start experimentation over that so once that smaller objective is achieved and it starts providing reliable results then you know it then we have to progressively build the solution around and on top of it so please keep in mind that err models take time to ensure and that is the reason why we say that AI is a journey from inception to production it's not something that can be achieved in a day or for that matter than the fourth night even so it has to be something that has to be invested over a period okay so with that out of the way let's get back to the sales forecasting topic now in this particular slide I'm going to be talking about sales forecasting insights of primarily there are two types so qualitative and quantitative so quantitative simply put is the gut feel or the foresight of the executive leadership that is developed over the years of working in the industry um basically their experience so for example if you are a dairy business and want to launch a new ice cream flavor in summer in the upcoming summer then in such situations you're mostly relying on the direction given by the executive leadership regarding the products launched its target market demographic the time on which its its launched and its sales targets of the other type is quantitative as the name suggests it relates to numbers more specifically historical sales numbers so since we are talking about sales forecasting its from a quantitative standpoint its carried out by using a methodology which is known as time series forecasting it's so let's talk a little bit about timeseriesforecasting in this lake or time series forecasting is used to predict the future events by analyzing the trends of the past I guess where it comes in line with AI so more specifically the data analysts or the data scientists examine the historical data to check the patterns or to identify the patterns such as trends seasonality side cycles and regularity and besides sales I mean sales is not the only department that can utilize the slit methodology other departments in the company like finance operations and marketing for example can also use the methodology for inventory and consumer demand purposes you all right so let's talk a little bit about the popular forecasting techniques so on a broader level we have two types or two types of two approaches in a way so the first one is like taking a statistical traditional statistical approach well these I mean stick if you were to go with statistical modeling there are a number of algorithms and models available and they have been in use for decades now they are still heavily in use um and I just mentioned you know don't know a few of the more popular ones out there but as I mentioned before there are quite a number of them that can actually be used during the exploratory phase by the data analyst or in order to identify the best or the most suitable algorithm that fits the given requirement so some of the more popular ones are auto regressive integrated moving average Oh algorithm which is a mouthful but it's you know the renown is ARIMA and the other one is simple and by the way I'm just going to be showing a demo of an experiment of an emo so in a bit and then we talked about simple exponential smoothing this is another one and the last one is whole windows exponential smoothing okay aside to the traditional statistical approaches you know in recent times another of course has also you know come to the fore which is machine learning this by the way is disrupting the way things that had recently been done and it's also seen as an opportunity to unlock some of the others other challenges also that were like seen as very difficult to find solutions for so as far as this webinar I'll show you an example of a neural network and I will also show give you an overview of the most powerful one of the most powerful AI services that is available in the answer for platform you ah just gonna stick me by it so since we talked about machine learning I think in this light let's talk a little bit about machine learning in little more detail so machine learning pipe primarily is a data science methodology that allows computers to to learn from the historical data to forecast or predict the future behaviors outcomes and trends so machine learning enables systems to learn without being explicitly program so I mean this is the biggest takeaway here when machine learning is compared with the traditional software development because traditional software development is more logic driven whereas machine learning is more data insight River so some of the examples some of the more popular examples that I'm just going to be talking a little bit about by the way there are a handful onp are quite a number of them but we talked a little bit about the one side I mention on this right here time series for testing which is by the way the topic of today's webinar there's one of the applications of AI the other one is IOT predictive analytics and an example of that would actually be an AI model predicting the remaining useful life of the industrial equipment and knowing ahead of time when it appeared is due or replaced product recommendations um example is like buying from Amazon and Amazon II that pretty much all of us have experience of going down so based on your buying is Phoebe AI or the machine learning algorithm makes recommendations for further impulsive purchase or odd buying or purchase them this has created a lot of value for the company by the Bursar's like a strategic initiative for Amazon ah another example is like fall detection so banks I have started using machine learning to flag transactions that do not see look right or more specifically at for donut so rather than having trained human analysts of flag systems axis which as you can imagine is quite challenging considering the sheer number of transactions committed in a day so machine learning is really good at picking up nuances and subtleties by the way and is ideally suited to augment the human workforce which is already being done okay now or moving on to another example is cancer detection so last and certainly not least if the massive improvement AI is bringing to healthcare so one such example is the detection of for running lives of lung cancer and machine learning success rated separately detecting the disease was 97 percent as for the study that was conducted by Nvidia a couple of years back right so moving on I'm just going to be showing you of a couple of experiments but before I do that chart if you have any questions being submitted so far then maybe we can pin them right now thanks Sam I do not have any questions right now from our audience but as soon as we do get a couple more I will let you know okay sounds good all right so right so let's talk about or the ARIMA model here this is an experiment which is a statistical experiment and if I were to show you the the data set yep so it's a very simple use case and in the data set we have two columns only one is like the paint dimension which is a frequency which which has a monthly frequency and the other one is basically the energy they consume this is a consumption call basically the number of items sold away so it's essentially telling us the number of items of a given product type being consumed in a month and the idea is to predict the the forecast over a period of time so so I mean this is like a univariate use case where you know we are only using one feature for prediction which is the time dimension and the data set does not have other features such as like weather patterns or demography or age groups etcetera that may or may not have an implication on the outcome here but as I said the for itself is a simplified user so we probably start with that and also give you a flavor of how it's it's done using you know pythons Tecna you know pythons data signs you know frameworks and stuff like that and and in and this whole solution is actually implemented in a Python ipython notebook or a two bit of notebook as it's being called these days okay so here in this step we have imported the data set I mean it's just going to show you the first five you know what rows here but you know it actually has a lot many rows essentially and the first thing that we do is we do a plotting of the input data so on a timescale so if you were to see it on x axis we have the other time times in spotted and on the y-axis we hansel we have the consumption which is the unit which is basically what you wanna predict going forward so this is currently how the data looks like and if you want to see from a visual standpoint which by the way is really important this this visualization techniques going forward especially if you have huge data sets available which go in GB so it's very difficult to eyeball these and editing and then gears relying on the result tooling that are available out of the box in the in the technology stance that that that you have chosen so another thing that I'm going now scale back to then an earlier point I made that you know that the choosing of a specific technology site is also very important anyway so if you were to see that this actually shows a pretty much linear kind of like a protection here of the number of items sold over over a number of months so it shows an increase in sales over time okay so then comes to the step of doing you know more pre-processing pre-processing here before we fit the ARIMA model on the data and one of the things that is the requirement here is to ensure that the stationarity is there in the in the data protection which which is basic which it basically means that the statistical properties of a time series they do not change over time and one way of knowing that is to see whether the mean is it's constant as well as the standard deviation mean by the way and sequency is also is not constant intensity point up with a pallet there so the idea is to pre-process this data bring into into a representation where it's actually more suitable to be to be provided as input to the ARIMA model and in the interest of time or not like going through each and every step here in detail but the ideal circuitry process this data we have taken for ARIMA ARIMA by the way is an amalgamation of different algorithms one is the error autoregressive models the other one is the moving averages mark so the idea is that we import that model and we provide the P process data seconds input to that and we try to fit that model on on on the on the pre process data and once that model the spirit the end objective of this experiment is to do to see as to how we are projecting the the forecast AF so if you were to see again we are doing some plots after we have applied the model here and we're letting the model make predictions for us so if we were to see this area which is highlighted right so this is basically your projection of the future sales forecast and and you can see that I mean we have stretched the time out in a way and over the period of time it's telling us that's going to continue on the d-mail and I mean as I mentioned before before that you know it's a simplified you say it in order to seem very practical at this point depending but the whole idea is to like an old take take a simple example because it's easier to explain as well but you know as for this experiment it can it is showing you with 95% confidence of the outcome of this model that you know the fourth cast is actually going to continue on the same line okay so the a with a reamer being explained we spice things up a little bit in the other experiment which is the machine learning experiment so we said okay it only had one dimension it only had the tiny dimension what if you add another column enter all right what if we either seasons column I then you see whether that helps us in any way shape or form or that has an impact on the on the way this model is going to or the model is going to behave or predictions are going to get affected by that so we started doing some more experimentation that was it and in this experiment we were thinking about using an uber network instead of using another statistical model so we change this data cycle little bit if you could see and clean this data column is the same as was being used in the in the experiment on your experiment you have the same target color which is the energy being consumed and then we have introduced another problem and we're calling it seasons just like give me any other experiment it's time to start slicing and dicing the data clean it out for example I'm just going to give you one example I think this season some you're looking at the seasons and winter spring these are strains and mathematical models or airing worlds or even statistical models they don't understand spin so they have to be encoding even number so that they can be fitted into the model so here in this step we have actually replaced the values with one hot encoding so we encoded the string values and then we started moving ahead we started doing more and more you know pre-processing there and then this is again the charting that the partner who charted again on the time axis as well as on the you know on the y axis we have the consumption so it also is showing us pretty much the exact same production as it did in the previous example because it's the same data set for at least those two those two columns again checking the stationarity here sorting the representation here we have actually you know calculated time banks here as well that you need to fit into into the model of Ghana we have actually come up with like twelve lines here now that now this is this is basically an interesting step that is where you're going to be pulling in the machine and this is where the machine learning you know processing starts so everything else tied to that step was all pre-processing and preparing the data to bring it to the stage okay so so how machine learning actually works in a nutshell so much in machine learning and if you if you are using a supervised learning approach here you know the data said incoming data set is actually or pre-processed data set is actually split into two now I'm not gonna say house but in two portions one portion the bigger one then I mean there is no hard and fast rule but sometimes you know it's it's advisable that you know you take one portion which is like 70% of the three persons data 78 the one the other portion is 30% and it could be like 25 75 anything that suits the experiment because this is a thing guiding figures and it takes a few addresses before the models are there lively train but generally speaking you have a bigger portion that you use for training the model and think of training a model as if you're like making a child is not any knowledge so you are letting the model over here this is the input and this is the expected output and then you keep on doing that this is the input this is the expected output and over a period of time you let the model train and get better at understanding what identifying things or another main things right so so the learning set is actually source as input for training and then the when you when when your model is trained it's time to see the era currency of the prediction that it's going to come out with so how to test that a simple approaches that you take back now this time around test it out that you stashed away somewhere else that was not sourced to the AI model because a a model is only trained on the training later not on the testing data that is your remaining portion that you did not use for training purposes and you already know the outcome of that because it's light from the same data set so this time round during testing you only let the AI or you only provide a I model the inputs and you will hold the outputs because this time around here like the air model make a prediction for you all right so when you only provide the input and the AI model only gives it and predicts the output for you you take the output of the air model and you compare that against the data that you already have available as part of your testing set if the prediction is somewhat near to what you're expecting then it's fine is not if it's theorists I mean if the prediction is way off then again it's time to go back to the drawing board and you have two people doing that or maybe you know change things around during encoding or maybe you know tuning hybrid parameters of the model or maybe you know play around with the data a little bit more retrain that and may being less than data somewhat more because the data that you had was not enough in a way so eventually it's a rigorous process and you when you get to a stage where your model starts giving you results that are more in line with what your expectation is another more accurate then it's time to take this more or use this model in a given setting where you can actually use it in an application or you know solution where it's like making predictions for you so so that said this is primarily what we are trying to do here in this step here as we near the lucrative training sites testing sites all right so and and and we are actually now in this in this step or in the following steps here pulling in a machine learning model which basically a noodle network of the type ls' TM like this long short-term memory it only has one dense layer and here we are fitting the model on the training data set we could say that this is the x-axis being both why is the D outcome and and the epoch is haunted epoch this by the way is a fancy name of the number of iterations so it's this the idea is that this model is going to trade on the given data set 100 times over and have three each titration the ideas I just going to get better so it's going to get better each and every term such a it's an every election or any pop so once this is done you know it's time to take the protections from the model so you take predictions here in this step and then you post process because we pre-processed a lot of information it's time to post process this information so that we can make the predictions to the same unit so that we can we can do the plotting and against what the prediction was versus what the input data was so it goes through a number of steps for pre-processing purposes and like if I were to cut to the chase here this is how it's going to look like so if you were to see the the projection here so so this this blue line is basically the actual projection of the actual data that I showed you earlier in this earlier in this experiment if you go here this is what it is whereas this this red this basic is basically the prediction of the model for the test data that we actually carved out from the input as well and we are seeing if you were to superimpose that on the excel data we can see if you see closely that it's not exactly like the same as the potato was it's a little off but it's not way off here so so if you want to keep on you know refining it maybe you can actually get this closer a little more also but here's the thing it's your projection because exactly like how it's supposed to be and it becomes exactly perfect then I times that that is also an indication that your model is over 30 so it may not always I mean it it doesn't begin it can never be hundred percent accurate it's going to be machine learning recommendations although the outcome is always going to be there and data ball point it has the closer it is to be absurd projection is it's a fighter but it it feels like exactly the same and then this an indication that something is not right something was not properly given training purposes anyway so it gives you a flavor of you know how you know you go about training machine learning model again it's again it's a simplified you know example because you know in maybe in real life you have like again lots of data where but it has a lot of features that that can be put into consideration and it goes through I mean you have to go through you know the time or or basically you have to go through the rigorous exercise of you know conducting the actual training of the whole model and this thing actually takes time but this is basically something which is very important because the moment you have a reliable model out that can actually be integrated across situations okay going back to the slide deck now that we have seen couple of experiments one was statistical model the other one was a neural network approach it's time to introduce a jorah order on them so as your order ml as you may have guessed since it has the name agile in it it's a Microsoft cloud AI service okay that helps expedite the whole process of identifying the most suitable model for a given useless okay so it's an attempt by Microsoft to help operationalize AI even further I mean it helps in the automation of time intensive tasks it rapidly hydrates over many combinations of algorithms and hyper parameters against your supplier data set and it helps in finding as I said before the best model on the success metric of your shoes and I'll show you that real quick in a moment or so accessible by a board designer and the SDKs SDK notebooks so I mean it has a LED UI also and it has also like the designer I mean that UI as well as it has the SDK so you can actually be used inside your Python notebooks also okay so so what is the so what is the detail so so the whole thing is that you know you read key auto ml take control because the whole idea is that you are going to be uploading your data set you're going to be uploading your data set to to the cloud service you are going to be configuring your cloud service properly and then the idea is that you know you like is your ml take control so and so the ml is going to take your data set based on your configurations is going to run on number of when I say a number I really mean it that was because I'm just going to be showing you that also but the idea is that you know it it runs multiple models again so data set and then after that it's done with your with with a cable run it's going to be coming up to the recommendation of the most suitable model for you that hey this is the best model for for your given useless and allowed me to explain what I mean by that so I go here yeah do you have a second question real quick okay but uh uh yeah this kind of was speaking to what you were talking about earlier in your time series model would you be able to do comparisons of measure forecast error like mean absolute deviation mean squared error so that but the question that we received from the audience and could you talk expand on that a little bit yeah I mean I think obviously yeah absolutely I mean they're they are actually a number of success metrics available that you can you know validate the outcome of your model against and see if your model was actually performing that's one of those statistical approaches or basically more scientific approach of seeing exactly how your murder was performing actually I'm just going to be giving an example of that using the answered ml you know demo or o or or and overview that I'm just going to be showing so I think it's meant to be answering this question in a little bit of more detail as well ok so although auto ml so rather than I just do the whole thing I'm just going to show you that you know you can you can configure that on the on the web UI and the I mean I'm just not going to be creating a data set I'm just gonna use a data set that was already used in one of the experiments before we can create a mutant here we can name it like test run okay so we can actually view data set also if you want to by the way it's the same data size so you have like the paid and the Energy Agency you also have the seasons here okay so Carter : for which is the prediction we're gonna say this is a target column it also has stupid you know specify a cluster because the cluster of VMs on which this whole training is going to be performed or they Cynthia stop running you need to be insecure they're not just gonna click on one of those here okay now in this step is this agile services giving me hints as to what is the other tasks that is that a classification problem that you're trying to solve it is this a regression problem but it's a time series forecasting table we're gonna think this is a time series for decibel and okay then tell me about your time dimension here I'm just going to have to start somewhere that there is that problem okay and it is here we can also like you know specify a group by columns I mean if assuming if your data set has a number of other features also that can impact the outcome you can keep choosing them here right and then you can hit finish unless not going to do that because it takes a while I mean even though on our data side it goes takes close to 20 25 minutes or so but I'm just going to be showing you the outcome of one of the rounds that we executed before so if you were to go to experiments here but click on that this was a run that was executed earlier so so the result is saying that you know the algorithm right so the algorithm that was most suitable for the given data set and the angry metric of success which was normalize would mean squared error and this is the the value that was basically the result of the test run here and so the for more further it is to zero the better it is so the lower it is the better it's performing your model in a way but you know there are other success matrices also that or pending latencies that we can specified I will give you an example is that also so those are the ones that are available we can by the way you stand in our Python both books as well if you are doing carrying out an experiment ourselves so like me I also do that and that was the question that was being asked or the normalized mean absolute error so we we can use or we can specify those here in auto and then also and we can also use those in in in in our experiments as well that we are going to be carrying work in which we are not using the capabilities of of the sort of woody ensemble was the algorithm name that was that was the best suited out of how many if we look at the number of models that were being used in it in the specific run I mean diesel runabout apart so Auto ml took in our data set and also configurations and start running that data sets later set in the configuration settings tools many those my meals bottoms right now we can also limit the number of algorithms because it's the if sub function of V of the the the input data size because the bigger the data size is the more time you start to take but the ideas we can also limit the number of X number of algorithms and also during on configurations and hey we don't want you to run it against each and every algorithm that you find but I just want you to run my data against only a handful of them and and tell me as to which one was the most suitable okay so would click on that and in and we see that if this is the one that we warned and maybe we can deploy that right away if you were to click on a deploy button here that model goes and becomes a web service so it becomes like an endpoint and you can actually start using that in your application in whichever I mean depending upon your your requirements in a way I mean could be a mobile application cookie another application or you can actually download that as a picker find out some yeah so you download that as a pickle file you can import that in your Jupiter notebook and you keep on going with your experimentation that places that so I mean it it helps expediting the process a lot take a look at it from that perspective so point back to the presentation size Spira metric film for my site if you have any questions please feel free to eat less and with that I'm just going to hand it over to touch thank you Sam really really informative if you could could you go to the next slide for me okay well if you would like to learn more about using artificial intelligence machine learning to help forecast sale we are offering an hour-long workshop we'll brainstorm with you discuss your sales forecast acknowledges other aspects of your business that might benefit from utilizing a on for example we recently have built an AI based invoice automation tool for accounts payable and we built a sentiment analysis tool to help gauge customer satisfaction at the end of this workshop you'll have an AI adoption strategy that you can utilize for your business as for the goal of our webinar today we'd like you to have a better understanding about AI and that is AI isn't scary at all it's just another business technology you can use to improve your sales or other aspects of your business the bottom line is embrace not for your AI we will be sending a copy of our slide deck a recording of today's webinar and a link to our websites AI page this is where you can learn more about using AI for your business on next slide Sam in the end if you do want to get a hold of us you can use this email address info at alpha bold or any one of these phone numbers to reach us directly again we will be sending out a email with a copy of the webinar and links to you where you will be able to get more information about AI and your business and respond to set up that workshop and I guess that about wraps up our webinar today are there any other questions yeah so check there is one question sure Sam the question is how you import the data set in Azure and you please show that oh here you go to a short order element you click on this automatic done you create a dataset then from here you can specify the option I mean it can be from a web file it could be from a data store and could also be a local file so you click on that and you browse to your data set I don't have one data set available on this system right now but I mean if you have like a CSV file or any other form in which you have the data set that will be choose one of those options and then it's going to get up go to to to be a SRS just like we did with this this one and then there's another question as well what's your favorite tool or ya boi casting the iPhone at the are and your Oh probably I'll also comment on that oh right so I mean this is a very interesting question I mean this is like pitting one formula against the other I mean you're talking about science here it all depends upon exactly the I mean see the whole the whole the the beat of the whole AI process is basically Queen and doing an experiment and that is why this is not known as developing an app this is actually thrown is doing an experiment and I mean this is very different from India to software development so in a way I mean whichever strategy seems or which I mean you you have to employ you have to make your model go through a number of models and to see the outcome of that and that is exactly what how to a man is trying to automate for us but the whole idea is that you know it depends upon exactly what you're you stresses in some cases if the prospects or the customers a lot as no customers and that's fine we can actually help them implement you know AI Basu business using open source technologies as well most of them are like pipe are available in the Python ecosystem you know so we have experience of working on tensorflow or to implement neural networks we also have experience of working on you know statistical models using libraries like the one that I showed you for Lima as well as for scikit-learn that is available again in in Python and you know this is this whole area this is becoming more and more consumable over the period of time because previously it was just in the realm of the data scientists and these guys are not developers but over the years now there is a shift in that trend and it's becoming more and more developer-friendly so you do have a java javascript you know based frameworks available novice well you have attention full case you have you know you know other things that are available especially in the Google platform in which they can they enable you to carry out experimentation on the client which actually could mean your browser or your local machine or maybe in some cases your mobile you know your MA mocha devices as well so I mean it so so the answer to this question is it depends on upon exactly the kind of problem that you end up trying to solve an away and just two after that we might find some of the experiments that are being that are already available in different formats in Python so you would take in that information you've probably put it in IR and then you would see how that performs versus some of the other models that are available so it's probably a combination of all of them as well to some extent that you are used to be able to come up with the best solution for your a problem right I'm glad you brought this point of just just a couple of quick words on diet I mean transfer learning is actually becoming really popular it depends on exactly what this is which is really popular especially for image classification use cases so rather than you start everything from scratch or you actually you know use some of the models that are already built with the image net which just train on billions of images already we just you know give it the pictures that you have available and then you tend to reuse the learning that is already built into that model and try to extend that learning to your viewer picture so that is becoming really popular and lastly Microsoft is is Microsoft supports all those frameworks by the way it's one just like they explain that you know it's it fully supports technology slight chance of loss so that's no good yet I don't have any other questions well it's perfect you know thank you everybody and just for a little context as far as you know hey or am I the right size to use you know AI or machine learning to improve my sales forecasting we work from with everybody from fortune 100 companies all the way down to startups to help them with their AI for sales using AI for sales forecasting so we specifically have targeted medical device companies and we have a couple companies that we used to help them with forecasting and like I said even small startups we work with a medical headlamp developer that's just getting out of the gates and they're using AI so we do have you know whole gamut of people that we do work with both pharma karmasu tackles medical device and even energy companies just to give you an idea of a couple of the industries that we work with where we've been able to help them using AI algorithms to improve their business so on that note again we will be sending out a copy of the webinar a link to our webpage our our website's AI page and the slide deck from today's presentation if you do have any questions feel free to reach out to us and until we talk again have a great day thank you all I think there everyone thank you you