ClearML and Aporia Partnership – Liran Hason, Aporia & Moses Guttmann, ClearML

This is Textron TV. Hey everyone, welcome back to Tech strong TV. I got another great interview lined up here.

We do a conversation. I have two CEOs both coming to us today from respective offices in Tel Aviv. And we're really happy to have them.

Let me first introduce you to LaRon Hassan Lauren is CEO founder of ahoria. Hey, Lauren, welcome to Tech strong TV. Alan thank you for having me my pleasure before we introduce Moses our second guest.

Let me give a chance give people a little bit of your background Laurent if you wouldn't mind and a little bit of the Emporia story. Awesome. Yeah for sure.

So on to see you with Emporium. Aporia is a full stack observability platform for machine Learning Systems, which means really from the moment the models get through production. Apoya is there to help you really get clear visibility To what decision the model is making what's impact does it make to the business guests live alerts on data drift performance degradation maybe on opportunity to improve your model.

So really everything that happens post-production so you can be on top of it. That's it. That is what we're doing in Emporia a bit about myself.

I come for over 20 years and software engineering being an architect of machine Learning System really from training to serving all the fun stuff with getting ml into production. Excellent. Thank you.

Our second guest is Moses Goodman and and Moses is the co-founder CEO of clear. Ml. Correct pleasure to get it right?

Yeah, right perfect to have you Moses same to you. If you wouldn't mind give us a little bit of your background and the clear ml story. So clear Mal actually started as an internal tool but like the cliche we open it to the world, but we also open source it to the world.

So we are one of the only machine learning operations and to end open source solution out there really open source client and website UI the back and everything. And what we did is we internally that was basically a spin-off of a previous company. We built an entire infrastructure on how to operationalize machine learning.

So think of it as bridging devops and machine learning basically how you do you automate this process of development starting from development and then bridging that Gap from running it on my own machine everything works to everything breaks on the cloud or when you deploy it. And the idea is always the ICD. So basically we build the system from the grounds up to support this motion.

So you can constantly automate this process even while you're developing. And that makes it yeah I get makes sense. Excellent.

So so guys look the whole area of machine learning is one that you know was overhyped to begin with I think and then you know got confused with AI it still does get confused with AI and and now if I think finally starting to see real machine learning kind of benefits and you know successful applications coming to Market that people can can use one of the areas where this was true is ml Ops. Right. I first heard ml Ops.

First on the scene. I don't know four or five years ago. It was sort of taking the place of what was application Performance Management APM and and stuff like this and then, you know quite rapidly actually observability.

Came up, you know. And and kind of over it took everything right everything's about observability now. And I don't know if you would classify as mlopsis part of observability.

It's different than observability. They their partners. You know what Lauren it sounds like a warrior is about observability and ml how would you and then Moses?

I'd love to hear your take on it right after but what's the relationship here? True, so that's a good point. I think that MLS really started a few years ago as a term, right but in the recent two three years, it really became a huge market, right?

So what is MLS the same way that phosphora developers need to have proper infrastructure proper tooling proper practices, right? And therefore you have a whole Market called devops with devops tools and devops Engineers as the ml executive system have evolved in the last few years. This need also became very significant for machine learning practitioners.

And that's how am a lot became instilled off its own with dedicated tools like clear email on the pooria. Right and we see more and more companies adopting these kind of cold back to your questions. What is the relationship between MLS and observability?

At least the way I look at it mlops is a broad Market with different tooling and I think in the last few years start to get a shape of different categories. Ml platforms like clear email is one category. Observability tooling is another category ML observability and that's kind of associated.

Say Json Market as data science ml teams when thinking about MLS, you should definitely think about How do you build your stack which of these tools do you need them to evaluate accordingly? So that's a kind of my two cents on analogs and observability. And maybe I'm maybe recently I noticed that for a lot of people when they say Emma Ops they actually mean NL stack.

And this is a really important point to stress here when we think about ml Ops. It's about the workflow the operation not just the point solutions that are kind of the ml stack if you will. and mlops at least the way that we think of it and and definitely in this partnership it really highlights.

The value of that is when you tie up all the different point Solutions into a workflow. And that workflow brings you value. So you have to build your own ml stack internally your point solution and whatnot.

But the actual value is having everything. orchestrated together playing together and then you really realize a true potential of your machine learning models because at the end this is an iterative process, right? You cannot expect first time you deploy something it works the first time you test your software obviously breaks your QA teams will tell you.

Hey, you have a bug that's not actually how it works with machine learning. Actually the the first time you'll really get a feedback from the field is when you deploy into your production, hopefully net Canary set up so you don't break everything right? And that's exactly when you need a tool like caporea to actually give you some feedback.

Oh, this is not exactly what we wanted. But you want to that feedback to be straight but to be to go straight for into the development process so they can very quickly iterate over the next version of the model see what went wrong and deploy that one because it does take a few iterations until you reach that point where your machine learning model and the end. Deployed and the bottom line is it needs to make you money.

I mean you're gonna invest a lot of time and effort and resources into that. And any and at the end you have to see the return on your investment. A great I I don't.

I think that's good like the both of you we put that together and I think are already inside something there. So I mean look guys. I have yarn here today because recent announcement or partnership between Emporia and and clear ml.

I don't care which one you actually we let Lauren go first last time. Let's let Moses go first this time. Moses that what what's this partnership about?

So this partnership so first neuron is an awesome guy. We've known each other before a while. So it's always great to work together and we were talking and we we kind of realized that we have a lot of common customers actually using our products in Tandem and we thought okay, that's we should probably make Fair life easier by better integrating and the solution itself Works quite well because the way that we work is clearly brings you from really from writing your initial code base to deploying your model on the cloud.

And the prayer brings that visibility on top and then feeds back everything into the development cycle. So you really end with an end to end flow a tie ties in the development units the R&D process and the business and the product with visibility to the actual performance of the model in real life. Yeah, so yeah, so I 100% agree with Moses and you know, I think when you think about it to our customers or ml teams and data science teams.

When they think about starting to construct their ml pipeline, right, they want to accelerate the speed of releasing new model to production. They want to have this pipeline from experiment tracking packaging deployment serving. This is where the clear mail part kind of provide you with that and then once you deploy to production, you run it as part of clear mill, then you want to see what outcomes are happening.

Thanks to your model when these the right time to re-train your model, right and maybe you want to automated process which he gets all of this report. So really by tying up is two solutions together. Ml teams can now get a full stack end-to-end poor ml platform.

So this is like the True Value we see our existing clients get as well as some new ones. I love it. Sounds like a perfect devops love story.

You know with this go all the way down to deployment then Ops and post deploy feedback loop make it better again. You know, there's there's a great story there. So.

When we say work together though, there's work together in this work together. What do you really mean? Hereby work together?

You know having beer here and there. Yeah that that's what I'm talking about. Exactly.

Drinking alcohol and gossiping. That's what we do. And then thank you so far time to time making sure that the products themselves really work.

Well with one another so you don't have to so the entire idea is you don't have to stitch things together. Yourself, right? It's out of the box there with everything just working and to end and that's a true value.

Otherwise, it's on the customers. And to do the stitching itself, that's exactly what we're trying to facilitate here with this partnership. So if I for instance though, is it API or just correct?

It's basically exactly so it's basically you have the claremual serving that does again from the development. Through model repository through deployment for with CLI and cicd processes. And then the serving itself basically report transparently.

Into aporia in a way that you can in real time add more metrics into the reporting itself and have everything visible into Emporia and then in Emporia, you can basically point back into callbacks. Okay? Why do I do if I have an alert when something happens?

actually, yeah, and and I think that the very Inception of this partnership really started just like kind of Late night conversation between Moses and myself talking about listen like we met with this view accounts earlier this week and they currently are interested in monitoring, but they also have some issues and challenges when it comes to tracking their experiments and how to facilitate all these pipelines to serving. Do you have any recommendation? Right and I know clear email has great product I said, you know Are you guys interested and apparently Moses and clear email had kind of?

A similar but the opposite question getting from their clients right about now. Also, I'm clear now. We enjoy using it we get the full stack until serving.

But how do we get visibility? How can we track and get alerts to slack Microsoft teams on theater drifts, right? How can we share other stakeholders with what what our models are actually doing and we say, you know what, let's maybe find a use case that we can collaborate on and once we saw it repeating we would decide you know, maybe we should just join forces invest in both sides to make this integration easier for users.

So it becomes just a few clicks. very cool Let's talk about the user experience then so. You know, I'm an organization that may be uses Emporia and have clear.

Ml. How do I how do I make the magic happen here how you know is just an automatic out of the box thing or basically, you've replaced your Helm chart with a different one and you're good to go it. Is that simple really?

Yep. The this is that that's basically the essence of the partnership. Basically, you have your regular setup your replaces with a new one and it just works you plug in your API keys and that's it.

That sounds easy enough it is. So, you know as they say in Las Vegas, this is a fine beginning Where Do We Go From Here guys? We have to take over the world like we do every night.

Okay, but you know, I mean, is there other areas that where we can you know, where where these two? Solutions working together Right, we can leverage that to do more. I think that the value itself.

of operationally operationalizing machine learning is really still a question mark for a lot of companies. So a lot of companies are still the Initial stage of adoption I agree, right? So at the end they think of it basically they're There are two main reasons to approach machine learning.

one you build an entire feature based on that in your product and you believe that feature will sell your product better or differentiate that from other competitors And the other is your optimizing your Revenue, that's the bottom line. So things like churn detection and predictionz and recommendation. These are all optimization of a product that it at the end is very close to the bottom line, right?

And in both cases, it's a lot of resources that you have to invest. Into adopting machine learning because it's very different from everything else that you're developing right? It's non-linear.

It requires a lot of people speaking very different languages and it requires involving the business unit and the product people that sometimes get their intuition and what can be done. From the media which sometimes is a bit more sci-fi than reality, right? And we have to communicate the capabilities inside the company itself, which basically means trial and error.

And because before that everything is on paper and this is exactly how you develop models for for months and months without getting everything into production and realizing if they actually work in your product and the product itself constantly evolves. And the idea of operationalizing this entire process is really the ci/cd. Concept that is relatively new right it's the last decade but before that it was only a few companies actually adopted that.

But it's the realization that you have to combine the two and make sure that this process works end to end. So the decision makers have visibility into what's going on and that visibility creates more iterations. And that's actually when you see the value itself.

It's not about making people more productive making your engineers more productive. It's actually seeing something working for the first time and understanding how to Optimize that for your own specific use case and that's exactly what you need that end to end. Solutions that bring you firstly visibility actually to see if some graphs and CEO.

It didn't work as I expected or assimilate. Oh that this is what would have recommended probably wrong. And actually feeding back to the R&D teams developing the next iteration.

And that's the True Value actually bringing those models more models into production and acting into actual use cases. And not to mention. They do change so data change product change, it's it's a constant revisioning of those models, but the first step is actually to get them to work with the way that you imagine your potential.

make sense Lauren I'm gonna give you last word because we're about at a time. About what sorry I said I'm gonna give you the last word. We're almost at a time anything else you want to say to wrap up.

Yeah, sure. So. You know, I think this is a great example for collaboration within DML space.

That really you know, the space is just in its early stages days yet. I think it's really important. We the problems we're dealing with the challenges are very very difficult.

And it's great that there are teams that are highly focused on each and every challenge along the way I think collaboration between different organizations, whether it commercial companies or nonprofit or just different ML and data science team. This takes all of us as the market as society that AI really affect all of us. He takes us to the next level.

So I think it's really important. This is just I think one great example of partnering up with another friends in the industry, but I really like to see more Partnerships more collaboration within our space. Hopefully hopefully we will hey gentlemen, I want to thank you for coming up here on text on TV.

And telling us telling our sharing with our audience about your collaboration, you know, we didn't mention people websites. com? com.

Yeah. And and Moses clear ml. Clear dot ml very clear dot.

Ml. Maybe one last thing I would emphasize is who would benefit from actually trying to find the right solutions for them in the envelop space and I think this is important because I just had a call today with a friend and he literally said I feel it's too early for me and I asked why and he said so because we're only a team we're of four or five people and it seems like Emma lobs even devops is something that you do when you have a larger team. But in reality and definitely in this economy.

Even smarter teams and especially smaller teams would benefit even more. From operationalizing and automating this process. It means that they are spending less time in basically infrastructure suffering infrastructure and Cloud infrastructure instead of letting the ml Ops stack or the mlops solution do the automation for them.

So they can concentrate on on actually buildings a model and testing it inside their product instead of scaffolding infrastructure. Just to test it. And it's surprising that this is kind of the approach but it's it's the opposite and definitely in this economy when we see companies not expanding so quickly.

Sometimes even the opposite we actually have to make sure that we empower the machine learning teams and the devops and then Machining Engineers so they can produce more with less. Absolutely, you know we're going through a similar thing when I first launched devops calm people was in an Enterprise thing. Is it something for startups that you have to have a certain size?

And and it worked it worked its way on. I I think it will here too. But guys, thank you, both again.

We're gonna break here our Tech strong. Thank you. We'll be back in just a moment.