Audioburst Search

Olivier Blais talks about the approaches to deploying AI that delivers the highest level of success

Automatic TRANSCRIPT

Aw I welcome to the pulse. Ai I'm your host Jason Stoughton here in Silicon Valley as many of you know I recently attended the open data. It assigns conference as a media partner and in this exclusive. PODCAST I sit down with a BA Blais as the CO founder of movie. I A high impact a consulting firm based in Montreal. I started by asking him to talk a little bit about move. Ai What it does and the types of clients at works with Yet differently So movie I worry Data Science Consulting Company. But we're really applied Redo applied machine learning so in other words we held to democratize the the the technology were based in Montreal. And you might know that Montreal is very strong technologically. Speaking with Yoshua Been Joanne Elementary. But we feel that mantra is good technically but were very bad at implementing concrete projects. So that's why we decided to act there and so what we do at depending on the metre three deep but we help to put in place a data science structure so that the companies are able to to start using this technology. We start building dignity stack internally we then at L. to build momentum so work on projects but in in a really applied the way. So we're using a lot of agility admitted and were or goal is to be able to deploy bike messenger projects very quickly to get vying from the business and like I said a momentum going. Is there a typical type of company. Work within in healthcare or retail. Or do you sort of work across the board. It's a good question we're asking yourself the question. Russian were a little bit more than a years old at right now were working with dot com with banks in manufacturing so right now of don't really add ethical customer but we often as difficul use cases because a lot of use case stink different. But then when you start thinking about the problem you're trying to solve it then becomes really similar to another tragic in another another industry so we still keep or is often and we're still working with a different Different types of industries. And as long as we're able able to find the right solution for the right prime were were okay with that and do you. Typically work with companies that already have data science applied machine learning teams. Or are they bring you in as sort of a new group of people and also who do you work. Within the organization you work with the CIO CTO CTO IT directors. What does that look like that? That's a super good question I thought that we could. We would just start working with companies without any team in place but then we realized that we were able to get traction in companies where the big team team but the D. need a little bit of L. often will We'll see the tenancy of having a team but too much too many projects for team because you're gaining a lot of traction when you start getting a team in place before you really know what so you have in in your backlog until you get team then you get the team used a okay. We could do all of those things. You always ask a bigger backlog than the team. Then the size of your team assault were helping teams to to deliver and also were very strong on on the Muddle Edition version In fact I presented on At the open the France so obviously I presented It's a thorough on muddle vegetation as so when there's a big team in place what we We size that. There's a need for the team to provided the model because often de model are seen good but not great and the dog Mr Yield take a good benefit for the forgive business unit as this is why we decided to act on that front. That's interesting which leads into my next this question on when you talk about model validation which I think will be a part of it but as you know ominous we all know in the industry you know the the the stats let's In the percentage of companies that are attempting on one hand to develop and deploy Ai Machine learning products. I'm projects compared to those who are doing it successfully. It's a stark contrast rise. Something like eighty five percent of CEO's according Mackenzie thank the machine learning is extremely important to the future of the company and Twenty nine percent of them have been unable to deploy any AI projects whatsoever and so my question is why. Why do you think that is what and what what they've been doing wrong? What are people who've done? It successfully doing right unite bullets from insights from young. It's a it's an interesting being a metric that you're bringing to the discussion because this is the exact same Exact same statistic that we're using and and it really impacted the way we we have shaped we reshaped or business. We ask ourselves the question. What can we do like I said? Before or goal is to democratize meshing earning and when we saw this statistics. It's all good What can we do to make? This is better so what we've seen so far there's the yield AC- sweet wanted to to ask but this lack of clarity because of the buzzwords at is for is one of the costs so so you're you're building team building alive but without any concrete projects to work on so it might be when the element. You're trying to too many things at the same time. There's no clarity. There's no Without any budgets or any projects defined vines. It's hard to find. Who should you interact with into the business units so this is a yeah? This is the one big element. Add another big element. Is the the the project mode. It's Lexi you know it's that science so there's a lot of the science in in this So we we have the tendency to work on projects like they were science projects projects but in a business context. It's very hard to yield benefit using this mode and I I would personally. I now realize that. So many technologies are available. Are there's a big community working on denser. Fool making it that volleyball working on a lot of very very useful technologies and Annesley. If you're one of your opinion a lot of projects it can be done in a very good without a big science background. End Let's thing that you don't need mathematics. You don't need to know linear algebra arrest at You need to have this but to work in the knicks per month even mode or the famous rouva consent is often very very dangerous. You work for. I don't know six months eight months and then at the end. You're you're presenting to your the team. Yes it's possible. It will take a year to execute the project when he could have executed if in the In a fast past week could have been done if you didn't go in this type of mode so this is another another problem in my Mike and this is why we decided to go to work in jazz modes at its heart but We've been able to demonstrate that the work's being able to great e backlogged and then deliver value. Every two weeks and often relies is a project that can that seemed third big can be a get can be executed in face in phases and deliver value. The I I don't know the first month or for two months after I started to work on the project. So you're you're denny boats who create momentum and also you're able to absorb any it changes from from the end users because it's new technology will always happen people change and the project will change and finally one of the big project were working on like. I said it's muddled volition. It's one of the most frustrating moment woman I think for any data sciences. It's when you're working on the project and it happened to anybody but there's a lot of data scientists who learned by themselves or they're like many data scientists have backgrounds and sometime you think. Did you have like a very solid solution but it's hard to have the right framework to validate that your model really works and the danger is seeing you know. This is a super good solution but then when you show if you present the final product that probably months and privy big team to work on and then the is wrong so that's a lot of frustration and you spend a lot of time working on so so this is why we think that a good Zayd Framework is something. That's important to protect duke to to to to protect your Your scope and also to announce the adoption to avoid this failure to Dan no create e Extension between between the company and that offense community because the community the community. The the company doesn't believe deep in the community that community delivers anymore. And this is very dangerous you touched on a number of great points their va and it's something that we do here. Also I and with our parents firm the foghorn group is you know trying I to help business leaders Within the company understand sort of the unique features of artificial intelligence machine learning the difference between in data science. which as you say is you know science right which takes a different approach than other engineering projects where you get a scope and you start working towards towards a rubber ball right? There's there's an experimental famous in all a data science projects right and that's hard culturally for a lot of organizations to understand but with that right and with those two things sort of in play there that you threw out who one of the things we keep coming back to is is that I think some of these projects are failing ailing because the organization that is trying to implement them whatever enterprise right with Telco et Banca retail stops from whatever. It is culturally literally within their company are not approaching these right that it's not a technology problem. It's a culture sure management organization problem right that maybe the IT side of the business Isn't the right one to do it. Maybe it needs to be on the business snus side and I'd love your thoughts on that it's in my mind. It's change management problem. You you work. The project will touch on several different business. Units and the data can be stored Ordina database. That's owned by it. Or you know where operations states the end users might be in marketing and the integration to her system you know the final integration to people present the EH predictions is on by it the infrastructure team. And so do you interact with those three groups. It's it's hard. And when you approach it with lack ed it's very advanced technologically. What you're trying to so you need to be able to demonstrate something quickly so that the different the different stakeholders? I understand where you're going with this end the opportunity to quickly interact with the tool of you developed a first version. That's very imperfect and then with this quick quick MVP It's phthalate apart because the puck. Is You awesome very incomplete and will miss a lot of Persian like the final integration and you will miss interaction critical interaction with like the final So let's forget about the puck but you do it quicken. Vp and then descend VP. Gets get get gets really really tested. So you're able to see the hurdles in the business and with that in mind than younger son okay. Access to you're quoted data is generating AQUI will need to deal on the infrastructure side this way. We need to find another role in the team to be able to better interact with the business units because we failed on the first difference version so enormous line By being Gile your vote to tackle on the the most important Element preventing developments team to deliver good value and this is management. That's that's a great insight. I'd I'd actually like to have more time out to another podcast on that in really really dig into that before I let you go. What are some your high level takeaways that you want our audience to know from that you got from od seat od SC? When the the bigger trend? That I saw during the odious conference is the emphasis on interpret the ability. I I used to recall two years ago last year at Several different people saying no. You don't really need to interpret. You don't really need to understand the black books if if the so if you need to understand the black butts means means you don't really believe in the project because I am able to demonstrate that ninety nine percent accuracy but yeah okay the that's good but you know in my mind. We need to go a little bit further than that. The goal is really to reach a doctor so this year we really aw is stepped in the right direction. A lot of people were talking about good and repeatability techniques approach and the were able to demonstrate yes it is important indeed to understand the type of friction but Nah just for the sake of understanding for disdain. The decision is direction. Understand the the blank spots of a solution so when it doesn't work Walleye is is it it why this is working and finally to be to give a little bit more insight to the end user. I saw it in my mind is one of the biggest trend that I've seen In the in the audience conference Well there you haven't. I hope you enjoyed the discussion. And don't forget. Follow me on twitter at the pulse survey and connect with them Lincoln until next time.

Coming up next