sfspark.org: Simon Chan Lightning Talk
Recording: sfspark.org: Simon Chan Lightning Talk
hello I'm Simon and you don't need to look at the slide you can just look at my jacket with a limited edition logo Einstein logo here so I actually mentioned I work at the cellphone Einstein group I'm reading the product management of the Einstein platform so if the magical word Einstein that everyone's hearing about that was Einstein so to explain it a little bit completely what it is Einstein every AI about sets what we call it Einstein but you can fit in two levels one is on a product level so we have different business lines right self cloud marketing cloud service cloud analytics commerce acquisitions last year so there are multiple business level lines and we want to add AI capability to all this product and so on a product level whenever is intelligent make use of AI machine learning deep learning technologies we call that Einstein but under that who is powering all this applications development so that's my group Einstein platform so there's a development platform underneath that enable all this groups the developers data scientists to build this intelligence application and we call that Einstein platform at least for now so myself a little bit background I started a company called prediction aisle about four or five years ago and the mission is very simple is to make machine learning AI development easy for developers and for data scientist and one problem that I try to solve in the previous startup as well as in cells was here is every prediction problem is unique how do we make it easy for developers to build AI applications you don't want to have a team of hundred people to build one think of predictive applications every time it doesn't make sense so it's their way that make it as easy as using a database server just standardized test analyze it so that everyone can learn the same skills and solve all the problems the reality is extremely complicated because back to the point every prediction problems unique think about it when you say the product recommendation engine which is the classic case of using data predict customer preferences you can say oh we just need one recommendation engine and it solves the world problem the reality is you see that company is investing millions and millions dollars to build custom recommendations solutions and it's because the business requirements is different the business go that they want to optimize is different the data is different users behavior is different in e-commerce companies selling closes it's very different from the insurance companies selling just a few products and you need to make recommendations so the first attempt I did is to open source the product and that's why prediction now is an open source project and now it's a patchy prediction IO essentially it's a solution for engineers developers if you want to modify the code if you want to add your custom business logic into it you can do it because it's open source so we provide templates easy to use but then you can modified it so the reason why I mentioned it is this approach works in a path because while if you're working in it right now if you're building single tenant application meaning if you're building solution for your own specific problem for your own company then it works now fast forward to the Salesforce context how can we can't just launch prediction out and call it Einstein why do we need a whole new solution the reason it's very simple because there are a hundred thousand customers in Salesforce and if we ask them to open use the open source create custom solution by themselves or maybe we hire 50 engineers for each tenon and then do it for them it doesn't make sense so the thing I want to highlight this in the Einstein platform the magic behind is actually multi-tenancy which is very unique in the industry when we look at existing solutions machine learning libraries the variability mentioned about the problem of motor tendencies when you're building applications let's say we build predictive lead scoring application in our mind we're not building it for one company in our mind we're building it for 100,000 companies which a totally different company they can be like five people shop using Salesforce they can be thousands of salespeople using Salesforce in the same day but then we creating a predictive lead scoring applications the same applications for all these different companies so what do we need Marc Benioff the CEO he has the famous closing behind every company there there are customers I modified it a little bit behind every AI magic there are data pipelines so what we need is a very production-ready data pipeline to get data in and we have a way to automatically build a lot of models so for each customers we need to build a lot of models automatically select the best one and deployed it in a very stable way so you can't say oh we deploy the model for 90 percent of the customers and we are very happy but how about the remaining 10 percent and when we do they designs evaluation like when you do modeling you need to do evaluations right be it precision we call but when we built one application we can't say oh it's great we improve the accuracy by 10% for 50% of the customers the rest of the 50% the accuracy job it doesn't matter we improve the accuracy so when we building the platform is very different how can we serve all these customers when we're doing data pooling when we're doing feature selection when we're doing feature engineering data cleaning when we're building models when we're selecting models when we're doing data evaluation all this because of the multi-tenancy we need to give it a new perspective and we're big user of open-source technologies Fox scholars actor and all this technology we also have got the contributors in open source world and as some of you guys know that like mr. Einstein is friend or well in business were required partner of mr. Watson so we work closely together in the open source space they have Apaches spot a system ml with IBM's box and ER and we support Apache friction IO with support Apache Linux so it's a very exciting space where we need a lot of production ready open source machine learning tools and I think that that's where mixes industry very exciting we're going beyond just building a model in a laboratory to production ready machine learning to production ready machine learning for a lot a lot of customers automatically thank you [Applause] you [Music]