sfspark.org: Edward Junprung Lightning Talk
Recording: sfspark.org: Edward Junprung Lightning Talk
[Music] [Applause] okay hi everyone so my name is Edward I'm the head of growth in sky mine and we're the company behind decline for J and the J Center Java and that's really important so my email is edward at sky mine do so if you have any questions after this talk feel free to email me so edward at sky mine do so some background we're founded three years ago 2014 we have about six million dollars in seed funding we graduated from Y Combinator we have about 12 14 mm clients about 4,600 people in our gator support channel locators where people from the community interact with our engineers 160,000 downloads of deep learning for J per month and we have about 25 people most of our deep learning engineers six have PhDs and our distributed across the globe because we only hire from our open source community and we have office in San Francisco and office in Tokyo so that's that's our background so what is timeline intuitively just think of us as red hats for AI or cloud era for deep learning is exactly the same business model and the one-liner is that we build custom deep learning solutions for fortune 2000 corporations and government using open source software so that's what we do and what we sell so classic open source nothing different services training and support so for services would you proof of concept we basically build a custom deep learning model tailored fit to your data for training we have two types of training public and private so public courses are just publicly learning courses and privates are private corporate seminars where they're both led by ascom an instructor and then support so we'll sign an SLA and we will ensure that Webber's build continues to work and this is a suite of deep learning tools all this open source Apache 2.0 which is free to use so keep learning for days our flagship that's our general purpose deep learning library that's where you build neural networks RL for JS for support reinforcement learning nd for J it's a computations behind be boring for J that's where we support GPUs arbiter is where we do model tuning so data avec is for ETL and model import is where you can import a pre train model from other popular frameworks like tensor flow so it's basically the bridge between data science and data engineering and now just zooming in on debugging for J and where the J stands for Java we don't do Python and that the three benefits is that it's open source Apache 2.0 and that basically means that it's free to use there's no licensing required and any solution that you build can be shared or resold with no limitations you don't have to pay skyline anything it's distributed so if you have like a large data set it could take months to train so neuron that neuron that training is very computationally expensive so you want this should be across a cluster of servers that's where GPUs come are very important it's almost required nowadays you'll have GPUs it's basically a joke and then it's written in Java and this is really important because big companies to use Java so it's data engineering and DevOps friendly and we integrate with widely used big data tools like Hadoop Spartan CAFTA which are all in the JVM and it's just a case study or ranch's it's one of the largest telcos in Europe we were helping common was helping them basically detect the type of fraud cult tells a sandbox tried using deep learning model called an autoencoder and basically what does so promise says so it tells the sandbox Fry's basically when someone hijacks an international call routed over voice over IP and injects the back to the network and that way to get a free call but it's a greater network quality for everyone and orange loses out on the fees so what we do is that we use an autoencoder and what we what a Dutch that it builds a profile of what normal call calls the bike and flags calls that are abnormal our significally different from the norm and in that way orange this analyst can focus only on calls that are obvious cases of fraud these are some of our other clients that we work with so arange telco fraud which I just mentioned trace 3 is working with a large aircraft manufactured suit for parts classification Daimler is for demand forecasting so predict the amount of cars to need to produce a couple months in advance Erickson was some kind of Q&A chat pod Homeland Security is exactly what you think basically it's travel risk assessment so what the next time you get on the aircraft basically they'll score you they'll give you a risk score to determine if you're a threat and then canonical which is basically a server failure prediction predicting which are Raziel go down and I guess the the common thing among all these people is that they all use Java and they all have so analogy they have so much data that they're human analysts cannot keep up so that's why they're interested a I and it's just our book Adam is our CTO and josh is our head of failed field engineering and it will be available at the end of this month and yeah this is just how we engage so basically for cuts the customer provide all the data let me show it your data and tell us what problem you're trying to solve like how you define success like what do you care about what outcome are you looking for and what's died mine provides is basically figuratively like this box which is basically the AI engine that gets you the result that you're looking for and I also show demo by I think I'll skip it we have a we have a workshop Friday at 10:00 a.m. and what we'll go through a demo and we'll go through how to set deep learning for J and now that's it you [Music]