Small Teams, Big Dreams: How You Can Get Started With Ai Today

Hi im carolina vesega, im, the chief product officer and co-founder at the australia. Ai im really excited about sharing with you, the presentation, a small teams, big dreams. How can you get started with ai today before starting this talk? Let me talk a bit about me and about study ai. I have a scientific background. I have been working in machine learning for 20 years now in different capacities from leading a research team at the university to working in applied matching learning, services and products.

In canada, a as a strategy, we help companies like yours to launch ai in your organization and the way we have to do it is through our sas automotive platform called kepler. Today we will start talking about what has changed in the last few years. That will bring us to the rise of automated machine learning, and after that we will be covering a fascinating real world example of how these small teams can apply machine learning right now so from there we will highlight some

Practical tips from generating value today and finally, we we will be concluding with some key takeaways. So what has changed – and this is not your typical, like ai history diagram, you know it usually starts in 1945 with alenterin. I i wanted to do something different. I want to actually talk about the last 10 years only to highlight the speed and importance of ai contributions made and and try to to to give some light on.

Where i think that we are going now from the business perspective, around 10 years ago, ai started going like mainstream in the research community that was held by the improvements in hardware and data. So now deep neural networks began to show capabilities, and after a couple of years, actually they even surpassed human capabilities in some cases so around. I think it was around 2015

Every company was hearing about ai and the hype start everybody wanted ai now in their company, but that wasnt that easy, so a couple of years passed and around 2017. In my opinion, that became the poc like large research projects. That was the norm. You need a research team, you need a group of phds if you want to do

Ai, because this is something very complex and actually the algorithms are complex and the data is complex and the programming is complex and and basically having a science phd. This was never better paid. This was like great for for scientists and but by 2018 people start to realize that there were. There were some pieces there like missing and it was common to hear that 85 of the ai projects never arrived to realization and this research and these big teams, with muscle learning operations and machine learning engineers, were very expensive and mainly only affordable by big corporations.

so quickly because the market adapts to everything 2018 2018 a lot of tools to increase the efficiency of data scientists start becoming more popular and the automated material frameworks and much learning operations tools ml obstacles start their race uh doing the market so now we are in 2021 and billion and science continue evolving and has evolved to the point where actually there are specific tasks that can be automated to the point where business analysts inventory planners business intelligence teams marketing analysts all these people who actually work with data in.

their day-to-day but that doesnt have the material knowledge they can start leveraging ai tools in their day-to-day by using these automated machine learning platforms explicitly created for non-much learning experts we have been hearing about no code look code low code a platform so what is automl for business and what can it do so here you have your typical data science flow and probably all of you know how this goes because you are all here so i i imagine you have already.

tried to launch data science projects in your organization but usually starts with a person who wants to solve a problem and usually this is either an optimization or is an automation or is a business process understanding and this person talks with a data scientist and maybe there’s there are some data engineers also involved can be internal can be external and after multiple iterations that often last for weeks or even months.

they can break down the problem they they can prepare all the required data they have built models they have they have optimized models now the next step is to prepare a report or a dashboard to start discussions with multiple stakeholders these discussions involve also additional information and multiple iterations that can happen so all these two steps process you have here is what we usually will call in creating this.

proof of concept and it might actually take months maybe depends on your primary three months six months nine months so and these are really bottlenecks for an organization because only the communication between the different stakeholders makes things slower and slower than what it should so the third step a step if we pass the first two is that now this model can actually not go to production in the way it was coded so now we need to make sure that the model is reliable that it can be implemented in an.

end-to-end lifecycle that has all the software development good standards that um that that we need and finally you usually need to integrate this model in your current tools either a website or a crm so business users or any other consumer can can get the benefits of the model so as you see this is uh this is long this requires multiple stakeholders and requires multiple back and forth factors and this is where autumn mail came to the picture so and and im going to explain you how it came first so.

there is certain steps that we can actually make automatic which ones from all this stuff you have seen which ones will depend on the tool there are different tools in the market and depends on who is the final user of the tool it might put more emphasis on one thing or another when its an optimal tool for business users you actually need to make the full end-to-end cycle automated while keeping and this is very important while keeping a human in the loop to.

make sure the business context is considered in the proposed solution so so this is the characteristics that the optimal tool will actually make all these process various moves is going to go from helping in the in the data preparation to deploying the model while keeping the business user in the loop not to ask much learning questions but to ask questions related with their business context so and what can be automated what are.

the type of tasks that can be automated and im going to make a point here the capabilities im going to be talking now and in fact what was achieved in the real example are all based on our experience with our own platform kepler uh so lets start talking a little bit about these uh these tasks that can be achieved the first thing is data classification so you for example you want to create a.

model that learns to classify users as turners or not turners or maybe you you your data entry is is a customer reviews and you actually what you want is to classify customer reviews according to the motions its an angry user its a happy user uh to make maybe what you want is to classify images imagine you are in manufacturing and you have a pipeline where you need to do visual.

inspection of your products so in that case you might want to have a model that is able to classify images in the damage or ready to ship additional classifying there are other tasks that we actually can automate a for example i can predict a quantity based on knowledge based on other variables for example a real estate agent might want to create a model able to predict the right the right price for a property in the current market on the market today or maybe somebody in an.

inventory planner who wants to predict a forecast lead time or forecast demand on a certain sq or pro there are other tasks that we can still do for example we can detect outliers or we can find patterns and segment things into groups that are formed organically by the system but in the in all these examples what they have in common is that even if there is not a unique solution to them there’s not its not that like there is one algorithm that will work well for every classification but what is true is that the science has.

advanced to a point where the machine can actually choose the right algorithm for you for your data for your case and not only the algorithm all the all the steps in the process so now now that we understand that things can go faster and we understand that there are tools that are able to do automated machine learning and automate the full flow for business users now i want to drive through this is a real.

world example it is based on an actual client and i have hidden some details for privacy reasons but the critical point that i want to make here is roommates so this is a company in the online reviews peer review reputation a is a website that serves both business users sorry businesses and consumers these consumers will write reviews about products and services and the businesses will address their concerns and get insightful feedback at the same time so you have the consumers and the businesses.

and this this is a mid-sized company with about 200 people a the team a that will that have had access to the tool that a is confirmed by two persons were going to call them one one is b d is a project manager and he works in the software development team hes passionate about automation find ways to give more value to the customers and we have v v is a product manager and hes in charge of business intelligence and hes the one who sets and monitors all the kpis in the business and he.

wants actually to understand what drives performance that’s his his goal both of them has a great understanding on their business and they they they have a big drive they want to excel and surpass any competitor so lets see in about a year they they actually finish five different projects if we actually he only has been working for two months with a platform most of the projects were were done by the both have other responsibilities in the business because this is a smaller organization with a lot of different.

aspects to cover but how do they start so the first motivation they had was to increase customer satisfaction and and this is why they start with the customer side but as they were understanding better what are the capabilities of the automobile platform they saw okay we can expand these projects to also other business areas and im going to go in detail a little bit on three of these projects the first one that i want to discuss is the user intent detection and this is super.

interesting so they knew the users go to their site their consumer im talking about the consumer the consumer go to the site with three primary intentions its either to write a review either to read answers to other peoples reviews or to read reviews to decide whether to buy that particular product or go to that particular store or not so knowing the user intention when he lands to the side actually was something very very desired by them because that would allow.

them to monetize and also engage further users so they also knew that this is a complex problem because you know this is not a property that is associated with one user so i can go to that site to write a review today but tomorrow i can go to read an answer or i can go because i want to buy another product and i want to know.

whats the opinion on that so so how can i actually understand the intention after i enter to the site and after maybe a couple of of clicks that i do in the size side the other thing that most of the users at the moment that they landed to the site they are actually not located so that makes things even more complicated because sometimes they can cookies or things like that they dont have a lot of information about what this user was doing before.

well they they actually a created a model is a classification model they wanted to classify the users in these three groups according to this intention and the data that they use they they had in their analytics uh platform they had the information about all the clicks that difference that are made in different sessions so they they create a data set where you had the first three clicks in every session and what was the last uh i the the last activity that they did so to understand what is the intention.

and they created a model that was eighty-six percent accurate they only took two to three weeks to actually a create this model and after that they were able to integrate that into their website make their website react to the prediction in real time and they they they currently do around 90 thousand calls to get predictions on user sessions every day so that that that was a very happy happy path and a very good story so after that the next step was again the with his passion of automation he said okay if i also can do classification he understood.

already classification if i can do classification but now with text i have so many customer reviews in here that need to be classified they need to put that according to the topic and they had an internal solution that they have created but was too as low the latency was too high so they decided to take the reviews that they had with the.

classification that was done either by their other solution or by humans and they put that in into into kepler into the platform and in only two days they they got a model and they were able to use the api i have a latest of milliseconds so they act and not only that the new model was nine percent more accurate than what they had before so.

look at this were talking about unknown material and expert that in two days yes he had the data that’s an important point but having the data in two days he was able to have the model have the model productized have the model in an api being able to deploy the api get milliseconds accuracy and a latency and improve the.

accuracy so this is pretty impressive but now at this point im going to talk now about the actual data revelation here at this point v b was our product manager and he was aware of the success of v because they were working a closely together so but he was facing a challenge and in his own a job and he needed the answers so and this is where i really think things became impressive one of the so what was the problem that.

we had he was mentioned in a kpi in their business and that kpi has dropped 10 from one month to the next month no clear explanation nobody knows whats going on here he needs to know whats going on so he he that hes a business person hes not a a programmer hes not a data scientist he has no machine learning experience.

but he has been exposed to all these discussions with me so he thought by himself okay give me that tool im going to take that platform and now im going to try to put all my data there im going to do a predictive model to see if i can understand this behavior and if i actually can using the modern interpretability and and the and the in the future impact if i can understand what are the key.

factors that are affecting this value and he was able to get a pretty accurate model he was able to find where was the difference between one month and the next month and he discovered that in fact this is super interesting that there was a variable that they were they were always collecting but they were not paying a lot of attention because that variable was always kind of stable but at the moment that that variable passes and.

threshold what is happening is that the kpi they were actually interested drops so now this helps a v to actually have a better business understanding he said this new this variable as a new kpi that hes monitoring and hes making sure he had initiatives in his business to to to ensure that that variable never changed again so this is pretty.

impressive and this is a window to the future its a window of what we expect that will happen when these tools come in the hands of business users so so since that the company has onboarded other four users and the they they continue improving with this these projects they daily they have ideas they they put that in the platform they trade they get some insights they use that in their business.

so whats the formula the formula for value generation what it is the first thing is ask the right question start simple so if you have a big problem that you want to solve dont try to solve all the problems at once think can i break down these in simpler sub problems can i break down this oh i i have a problem i dont know uh my my inventory planning is is being messy but i have that involves the the orderly.

time and and involves the demand forecasting and maybe involve other processes can i actually solve one get create a model for one item at a time so what go one step forward every time so the second thing is think about your data what data do you need and every time somebody asks me that my answer is always think like a human expert what information would you use if you as a human you need to answer this question.

in your expert knowledge what are the factors that might be influencing your answer after that think what data do we have and try to find what is the match between both of them and start if you have some data even if its incomplete you can start iterating you’re going to start saying okay let me see where where this data allows me to go and after that say okay you know these are.

the three factors more important start collecting more data including the data and and increase your accuracy as we go the the next step i think is important is is to have the right tool i need to be a machine learning tool for business users not for data scientists because you’re not a data scientist if you are in this capacity and there is something that me and now i have my my scientific.

hat on me is very important so this tool needs to help you understand when you can put a model into production its not only about accuracy there are multiple other things for example is is is my data in production going to be similar to my training data or is it going to model to fail or a is there any bias in my data is that it.

leakage in my data so make these are concepts that that probably you dont manage so you need to make sure that as part of the automation this tool actually has this gather rails and and you can trust you can trust this in size and you can trust that when you put this into production you are going to be served in the right way and remember start simple iterate and improve so my last slide is the key takeaways so i only want to say three things first is for specific standard tasks.

like classification prediction segmentation the best people to get predictive and prescriptive insights from the data are the domain experts you need you a person who doesnt know anything about supply chain this is very difficult for that person create a model about supply chain because there is key information that is required the second thing is that these tools.

actually they can augment the human capacity in your day-to-day actually the idea of these tools is to help you to do a better job and im for sure that the key factors are the data that is always important the tool and more than ever is important the writing and the right team is not only in the expert knowledge in the field the ability to understand your data but also in being eager to do something that goes beyond beyond the normal things that you do you actually want to excel you want to be.

the best so that’s that’s what it takes thank you very much its a pleasure i hope that you found this insightful.