0:03

My main mission here is to introduce

0:05

Muhammad. Uh so Muhammad Bedia is the

0:08

head of data and analytics at Canal Plus

0:11

and he's been doing some really

0:12

fantastic things with uh with Zerve in

0:15

his role and so let's give him a big

0:17

hand. Welcome.

0:23

Well, thank you. Um so I would like to

0:26

introduce you to our data transformation

0:28

journey. So exactly 18 months ago I was

0:31

hired by Canalas to start the data

0:33

transformation and uh this

0:36

transformation was motivated by the fact

0:38

that we used to be a satellite company

0:40

and we moved to internet and so we have

0:42

a lot of data points on customers and we

0:44

would like to move to more AI machine

0:46

learning and stuff like that because we

0:48

have realtime data now and and we have a

0:51

service offering like Netflix on video

0:53

video in demand and stuff like that and

0:54

also the fact that because of the AI

0:56

buzzword our sea level start hearing

0:59

about AI etc and you know as most of the

1:01

companies does this uh last two years

1:04

they wanted to tap into the potential of

1:06

AI so I was hired to to do this data

1:09

transformation but as as many of you

1:12

know garbage in garbage out this is the

1:14

fundamental concept in when you do data

1:16

analytics um I was in a kind of a

1:19

dilemma that most of the chief data

1:21

officer knows that either you spend time

1:24

strengthening your foundation putting in

1:26

place data governance putting data

1:28

engineering best practice and

1:31

strengthening your foundation to be able

1:32

to do accurate um uh data analytics and

1:35

AI and this will take you three four

1:38

years and without being able to deliver

1:40

you know fundamental and gamechanging um

1:43

uh product and stakeholders will lose

1:46

face on on on you and said so this is a

1:49

classic dilemma that um um chief data

1:51

officer are are facing so we decided to

1:54

have another approach um I will talk you

1:57

about this approach So to today we we we

2:01

moved from descriptive like the very

2:03

granular basic level of analytics

2:05

maturity data maturity to predictive and

2:08

by end of the year we will reach a

2:10

perspective and I will explain how we

2:12

reach that that level. So we decided

2:15

first to identify what kind of use case

2:17

we wanted to do with data what value we

2:19

want to create and so we focused on on

2:22

on these uh these pillars. So

2:24

acquisition we want to acquire more

2:26

customers, improve a customer

2:28

relationship, uh increase meant to

2:32

strengthen the relation uh retain

2:34

customers, develop content strategy, uh

2:36

cost-saving automation and do all of

2:39

this by uh being compliant with the eyes

2:41

of regulators. So we collected all of

2:43

these use cases to see what what are the

2:46

most important one of them and then we

2:48

put them into this impact versus effort

2:50

metrics. So to highlight to identify the

2:53

the the easy the quick wins and the

2:55

strategic project and then after we did

2:58

that we start we said okay we're not

3:01

going to uh strengthen the foundation in

3:03

one shot we will strengthen the

3:05

foundation within the data scope of this

3:08

each of these use case. So we've done

3:10

this basic uh basic data transformation

3:13

that you probably know. So we went from

3:16

a decentralized uh data space to um

3:20

snowflake and using AWS. So you really a

3:24

centralized repository we unified

3:27

business definitions and and put some

3:29

data dictionary and collaboration. And

3:31

this is did not take didn't take us too

3:33

much time because we were just focusing

3:35

on a very niche uh uh uh part part of

3:38

our our our data set. And then after we

3:41

did that, so now we're ready to tap into

3:43

machine learning and and generative AI

3:45

etc. within this scope again we wanted

3:48

to have a tool that will help us uh do

3:52

this because the data is ready. It's

3:54

clean. We have we document it very well

3:56

and I've I worked before in big fours in

3:59

the loit and PWC and I spent many years

4:01

in banking and I know that to put a data

4:04

science product into production it's

4:06

really really hard because building a PC

4:08

is one thing but going into production

4:10

you need to rely on it uh you need to

4:13

work with uh some you know sometimes

4:15

developers to put something in your

4:16

website etc. And I wanted to have a tool

4:19

that will allow me to put this uh use

4:22

case into production without this burden

4:24

because I am I'm not in the capacity to

4:26

go and talk to uh our board members and

4:29

tell them okay I need the I need to hire

4:31

X number of data scientists etc because

4:33

they have not seen value yet. So we need

4:36

to show some uh use cases some value

4:39

that then come back and ask for hiring

4:41

uh people. So this is where we relied on

4:45

Zerf. So I'm going to show you quick

4:46

some use case that we put quickly uh

4:49

thanks to to Zerf. So the first one was

4:52

um churn churn uh churn um churn

4:55

prediction model and this the

4:57

specificity of the churn prediction

4:59

model is that it's also

5:02

use explainable AI to understand why

5:04

customers are churning and then use

5:06

generative AI for our commercial team to

5:08

have a script and and talk to uh uh use

5:11

it by marketing to talk to our customers

5:13

to retain them. So we we follow these

5:15

five steps to understand why customer

5:18

are churning. Um we mathematically

5:20

modeled the churn phenomena. Uh we

5:23

selected a group of people to we would

5:26

like to act on and we build this. So all

5:28

of this has been built entirely inside

5:30

uh zer. So it is a a bunch of uh SQL and

5:34

and python with our uh codes and

5:36

everything was in zer. What I really

5:38

liked here is that with Zerve we we

5:41

connected directly to uh Snowflake which

5:43

is our data warehouse and you could see

5:45

engineers uh data scientists and IT

5:48

teams working into into the same uh

5:50

platform and this is this is quickly

5:53

some result of this model. So we it was

5:56

uh implemented into the Netherlands. So

5:58

I didn't mention that M7 span over nine

6:01

countries and Netherland being one of

6:03

one of the biggest market we have and so

6:05

we could see like a detection coverage

6:07

with this churn prediction model of 70%

6:10

uh when in the back testing phase and

6:12

when uh marketing team start testing it

6:14

we uh using different AB testing

6:17

approach we could see um uh uh good

6:20

results depends on the strategy of

6:22

communication you we did and this beta

6:24

testing approach was was used um uh was

6:27

was made thanks to generative AI.

6:31

So another topic which is I'm sure when

6:33

you watch something on Netflix on Prime

6:36

you you get instantly or the day after

6:39

an email showing you yeah hey you watch

6:41

this maybe you will look at look at this

6:44

and for us this is was like wow this is

6:47

very hard to to to use for the record

6:50

Netflix spent 12 years every year

6:52

spending 1 million on their

6:53

recommendation engine and we were like I

6:57

would say babies in the in in in the AI

6:59

world. So we wanted to have something

7:01

quickly. So we we builded this this

7:04

model inside uh inside Zerf using AWS

7:07

bedrock this this LLM. So everything was

7:10

compliant with our IT department and so

7:12

no red flags on it and these emails that

7:16

you see on the right side on the screen

7:17

has been generated automatically via API

7:20

calls within the observe environments

7:22

and it is running uh um real time. So

7:26

whenever someone watch uh and we have

7:29

like a buffer of couple of hours they

7:31

receive this these kind of emails and

7:34

the impact was could we we could see the

7:37

impact because as you can see here it's

7:39

just a sample of people that receive

7:41

those emails you could see a spike in

7:43

the engagement and this and and we have

7:46

we have demonstrated that also there is

7:48

a negative correlation between

7:49

engagement and and and and churn. So the

7:52

more you keep people engaged the less

7:54

they churn and this was a strategic for

7:56

us. Another use case here was the ch

7:59

search engine. So we were using very

8:01

classic search engine like um um elastic

8:05

search and here we wanted to use more um

8:09

uh robust search engine. So hybrid

8:11

search engine we we vectorized our

8:13

database. So we use a uh we used a

8:16

provider for that and then we tap into

8:18

bedrock within AWS and everything was

8:20

again done uh scripted into into Zerve

8:23

the a till the API deployment and so we

8:26

handed over this API to our IT and it

8:28

was like uh quickly to to to integrate

8:31

it into uh some of the app of certain

8:34

region and this unlocked something that

8:37

we didn't have in the past for example

8:38

like having movies when I describe a

8:41

scene I could I could have the movies

8:44

like a movie with the concept of blue,

8:46

blue pill, blue pill, red pill. This is

8:49

for for example was matrix or searching

8:51

for actors, producers and stuff like

8:54

that directly with within the search

8:58

and um another use case is the the the

9:02

way we do recommendation. So when you go

9:04

today uh in yeah those giants plat

9:07

streaming platform like prime and and

9:09

and Netflix you will see that the the

9:12

screen is really adapted the rows that

9:14

where you see recommendation if you are

9:16

in the same household you have four

9:17

profile the kids and and and and the

9:21

husband the wife etc. you will see the

9:23

different different uh profile different

9:25

rows and this is we were able to do this

9:28

within with with within Zerf and

9:30

something really important to mention

9:32

that I mentioned that we started this

9:34

journey 18 months ago so we spent like

9:38

nine nine months to 9 to 10 months to

9:41

strengthen the foundation and then we

9:43

started working with Zerve around

9:45

December January so and all of these use

9:48

cases we were using a local notebook etc

9:50

and then when we had we acquired Zerve,

9:53

we were uh were we were able to go

9:56

directly into production. Yeah, that so

9:58

that's it for me. Thank you.