Up next in 10
Retail is undergoing an AI-driven transformation — and developers are at the heart of it.
This session explores the developer toolkit powering retail’s AI revolution, from advanced APIs and MLOps pipelines to recommendation frameworks and real-time personalization engines.
Key insights covered:
• Retail AI research grew from 12 papers (2000) to 847+ (2023) — a 21.3% CAGR
• Modern AI systems process 150+ customer attributes in real time
• Transformer-based models like BERT4Rec deliver 23–31% performance gains
• Only 10% of AI implementations achieve ROI — we’ll uncover why
• Generative AI in inventory systems processing 20–50 features
•• Real-time personalization APIs achieving 75% accuracy
• Forecasting improvements of 30–50% with scalable architectures
Technical focus areas:
• MLOps pipelines for scalable retail AI
• API design patterns for recommendation engines
• Data preprocessing tools for customer attribute management
• Model deployment strategies for real-time personalization
• Monitoring frameworks for production AI systems
Whether you’re a developer, architect, or retail innovator, you’ll leave with actionable strategies to build production-ready AI systems that deliver business impact at scale.
----------------------------------------------------------------------------------
AI in retail, retail AI development, retail APIs, MLOps for retail, AI recommendation systems, BERT4Rec in retail, retail personalization AI, customer data AI tools, generative AI in retail, AI forecasting retail, AI deployment strategies, retail AI frameworks, AI-driven retail transformation
#RetailAI #AIDevTools #MLOps #Personalization #RecommendationSystems #GenerativeAI #RetailInnovation
Show More Show Less View Video Transcript
0:00
Hi everyone and welcome to a new episode
0:02
Hi everyone and welcome to a new episode
0:02
Hi everyone and welcome to a new episode of AI Dev Tools. In this episode, we're
0:04
of AI Dev Tools. In this episode, we're
0:04
of AI Dev Tools. In this episode, we're going to talk about DevTools API and
0:06
going to talk about DevTools API and
0:06
going to talk about DevTools API and implementation strategies on how you can
0:08
implementation strategies on how you can
0:08
implementation strategies on how you can build AIdriven retail applications. And
0:10
build AIdriven retail applications. And
0:10
build AIdriven retail applications. And joining us today is Mega who is uh
0:14
joining us today is Mega who is uh
0:14
joining us today is Mega who is uh senior program manager at Amazon. Hi
0:15
senior program manager at Amazon. Hi
0:15
senior program manager at Amazon. Hi Mega and welcome to the show.
0:18
Mega and welcome to the show.
0:18
Mega and welcome to the show. [Music]
0:23
Hey, thank you so much. First of all,
0:24
Hey, thank you so much. First of all,
0:24
Hey, thank you so much. First of all, I'm super excited to be here and thanks
0:26
I'm super excited to be here and thanks
0:26
I'm super excited to be here and thanks for this opportunity to talk to you all
0:28
for this opportunity to talk to you all
0:28
for this opportunity to talk to you all about my experience. Um, as you
0:31
about my experience. Um, as you
0:31
about my experience. Um, as you mentioned, I'm Mega Huli. I'm a senior
0:33
mentioned, I'm Mega Huli. I'm a senior
0:33
mentioned, I'm Mega Huli. I'm a senior program manager at Amazon. Now, uh, my
0:36
program manager at Amazon. Now, uh, my
0:36
program manager at Amazon. Now, uh, my background essentially is I'm an
0:38
background essentially is I'm an
0:38
background essentially is I'm an engineer by education and I have a
0:40
engineer by education and I have a
0:40
engineer by education and I have a master's in business analytics with
0:43
master's in business analytics with
0:43
master's in business analytics with University of Texas at Dallas. In my
0:46
University of Texas at Dallas. In my
0:46
University of Texas at Dallas. In my journey, I took very different roles,
0:48
journey, I took very different roles,
0:48
journey, I took very different roles, business program, product manager roles
0:51
business program, product manager roles
0:51
business program, product manager roles where very in-depth of technical, right?
0:53
where very in-depth of technical, right?
0:53
where very in-depth of technical, right? as a technical product manager or a
0:55
as a technical product manager or a
0:55
as a technical product manager or a project manager and worked across many
0:59
project manager and worked across many
0:59
project manager and worked across many industries like healthcare e e-commerce
1:02
industries like healthcare e e-commerce
1:02
industries like healthcare e e-commerce uh warehouse automations or you know
1:05
uh warehouse automations or you know
1:05
uh warehouse automations or you know fuel pricing and many more industries as
1:08
fuel pricing and many more industries as
1:08
fuel pricing and many more industries as I've tried everything to make sure that
1:11
I've tried everything to make sure that
1:11
I've tried everything to make sure that we are um in a AI trends or provide the
1:16
we are um in a AI trends or provide the
1:16
we are um in a AI trends or provide the technical knowledge that I can provide
1:18
technical knowledge that I can provide
1:18
technical knowledge that I can provide and for me the common thread in all of
1:21
and for me the common thread in all of
1:21
and for me the common thread in all of them is technology and data I have the
1:23
them is technology and data I have the
1:23
them is technology and data I have the power to transfer business into improve
1:26
power to transfer business into improve
1:26
power to transfer business into improve the customer experience.
1:28
the customer experience.
1:28
the customer experience. Uh because I always the person that who
1:30
Uh because I always the person that who
1:30
Uh because I always the person that who love tackling challenges and hands on on
1:33
love tackling challenges and hands on on
1:33
love tackling challenges and hands on on picking new technology quickly and
1:35
picking new technology quickly and
1:35
picking new technology quickly and finding a ways to improve the process.
1:38
finding a ways to improve the process.
1:38
finding a ways to improve the process. That's about me. Don't want to take more
1:40
That's about me. Don't want to take more
1:40
That's about me. Don't want to take more time on talking about myself. Let's get
1:42
time on talking about myself. Let's get
1:42
time on talking about myself. Let's get started. So um I have spent my last
1:46
started. So um I have spent my last
1:46
started. So um I have spent my last decade working on intersection of
1:48
decade working on intersection of
1:48
decade working on intersection of AIdriven SAS platform, enterprise
1:51
AIdriven SAS platform, enterprise
1:51
AIdriven SAS platform, enterprise automation and product innovation. Much
1:53
automation and product innovation. Much
1:53
automation and product innovation. Much of my work has focused enabling cross
1:56
of my work has focused enabling cross
1:56
of my work has focused enabling cross functional teams from data scientist to
2:00
functional teams from data scientist to
2:00
functional teams from data scientist to product engineers to scale intelligence
2:02
product engineers to scale intelligence
2:02
product engineers to scale intelligence system that just don't work and truly
2:05
system that just don't work and truly
2:05
system that just don't work and truly deliver value which is very much needed
2:07
deliver value which is very much needed
2:07
deliver value which is very much needed right now. Today I'm super excited to
2:10
right now. Today I'm super excited to
2:10
right now. Today I'm super excited to walk you through technical journey of
2:12
walk you through technical journey of
2:12
walk you through technical journey of smart dev assistant AI agent agent mode
2:14
smart dev assistant AI agent agent mode
2:14
smart dev assistant AI agent agent mode in VS code with focus on how developers
2:18
in VS code with focus on how developers
2:18
in VS code with focus on how developers can lead retail AI transformation using
2:21
can lead retail AI transformation using
2:21
can lead retail AI transformation using robust schools uh robust skills uh agile
2:26
robust schools uh robust skills uh agile
2:26
robust schools uh robust skills uh agile workflows and scalable infrastructure.
2:29
workflows and scalable infrastructure.
2:29
workflows and scalable infrastructure. Okay, whether you are building
2:31
Okay, whether you are building
2:31
Okay, whether you are building recommendation engineer engineers or
2:33
recommendation engineer engineers or
2:33
recommendation engineer engineers or monitoring models in production or just
2:35
monitoring models in production or just
2:35
monitoring models in production or just getting started with MLOps.
2:38
getting started with MLOps.
2:38
getting started with MLOps. Hope this session will help you uh with
2:40
Hope this session will help you uh with
2:40
Hope this session will help you uh with anything that you need. Let's begin by
2:43
anything that you need. Let's begin by
2:43
anything that you need. Let's begin by looking at the behind the
2:44
looking at the behind the
2:44
looking at the behind the transformation. Retail AI development
2:47
transformation. Retail AI development
2:47
transformation. Retail AI development has grown drastically rate. Right? If we
2:50
has grown drastically rate. Right? If we
2:50
has grown drastically rate. Right? If we rewind to the early 2000, as you can see
2:53
rewind to the early 2000, as you can see
2:53
rewind to the early 2000, as you can see in the slide, it's just 12 AI research
2:56
in the slide, it's just 12 AI research
2:56
in the slide, it's just 12 AI research articles that were published and it was
2:59
articles that were published and it was
2:59
articles that were published and it was f it was focused basically on
3:03
f it was focused basically on
3:03
f it was focused basically on uh decision trees mostly rulebased
3:05
uh decision trees mostly rulebased
3:05
uh decision trees mostly rulebased systems for things like if user buys X
3:09
systems for things like if user buys X
3:09
systems for things like if user buys X then pro give them the show them the Y
3:11
then pro give them the show them the Y
3:11
then pro give them the show them the Y all right or if the user buy A and B
3:14
all right or if the user buy A and B
3:14
all right or if the user buy A and B then maybe give them the C. uh we never
3:16
then maybe give them the C. uh we never
3:16
then maybe give them the C. uh we never thought about during that time that if
3:18
thought about during that time that if
3:18
thought about during that time that if the user don't know what actually wants
3:22
the user don't know what actually wants
3:22
the user don't know what actually wants it from the data that we previously had
3:24
it from the data that we previously had
3:24
it from the data that we previously had it to show it but today in 2023 we are
3:27
it to show it but today in 2023 we are
3:28
it to show it but today in 2023 we are in the era of transforms and re
3:31
in the era of transforms and re
3:31
in the era of transforms and re reinformation of learning and GANs these
3:34
reinformation of learning and GANs these
3:34
reinformation of learning and GANs these models power personalized experience
3:38
models power personalized experience
3:38
models power personalized experience dynamic pricing even supply chain
3:40
dynamic pricing even supply chain
3:40
dynamic pricing even supply chain optimization
3:42
optimization
3:42
optimization consider this number of AI publication
3:45
consider this number of AI publication
3:45
consider this number of AI publication in retail jumped from few hundreds to
3:49
in retail jumped from few hundreds to
3:49
in retail jumped from few hundreds to 847 in 2023
3:52
847 in 2023
3:52
847 in 2023 uh driven largely by ML tools like
3:55
uh driven largely by ML tools like
3:55
uh driven largely by ML tools like TensorFlow, PI torch and open cloud
3:58
TensorFlow, PI torch and open cloud
3:58
TensorFlow, PI torch and open cloud platforms.
4:00
platforms.
4:00
platforms. At retail comp company right being from
4:02
At retail comp company right being from
4:02
At retail comp company right being from there for example we use deep learning
4:06
there for example we use deep learning
4:06
there for example we use deep learning for everything from anticipating
4:08
for everything from anticipating
4:08
for everything from anticipating customer demand to autogenerated item
4:11
customer demand to autogenerated item
4:11
customer demand to autogenerated item description. These are just from the
4:14
description. These are just from the
4:14
description. These are just from the scratch surface from the data that
4:16
scratch surface from the data that
4:16
scratch surface from the data that seeing what our user actually wants it
4:19
seeing what our user actually wants it
4:19
seeing what our user actually wants it it from not from one user right it from
4:21
it from not from one user right it from
4:21
it from not from one user right it from the uh millions of users that has been
4:24
the uh millions of users that has been
4:24
the uh millions of users that has been used on those description and how do
4:26
used on those description and how do
4:26
used on those description and how do they proceed it
4:30
this search has brought into new kind of
4:33
this search has brought into new kind of
4:33
this search has brought into new kind of architecture right the modern retail AI
4:36
architecture right the modern retail AI
4:36
architecture right the modern retail AI systems are built on a foundation of
4:38
systems are built on a foundation of
4:38
systems are built on a foundation of realtime data and microservices Picture
4:42
realtime data and microservices Picture
4:42
realtime data and microservices Picture a system that analyzed 150 customer
4:46
a system that analyzed 150 customer
4:46
a system that analyzed 150 customer attributes in a real time things like
4:48
attributes in a real time things like
4:48
attributes in a real time things like preferred price point click uh clicking
4:51
preferred price point click uh clicking
4:51
preferred price point click uh clicking bound behavior or even device types
4:54
bound behavior or even device types
4:54
bound behavior or even device types right recent one of the examples that
4:56
right recent one of the examples that
4:56
right recent one of the examples that I've seen uh that right even if the
4:59
I've seen uh that right even if the
4:59
I've seen uh that right even if the people are buying something from Android
5:02
people are buying something from Android
5:02
people are buying something from Android the prices were very change compared to
5:04
the prices were very change compared to
5:04
the prices were very change compared to the Apple devices same way there's a
5:08
the Apple devices same way there's a
5:08
the Apple devices same way there's a trend that has been observed from our
5:09
trend that has been observed from our
5:10
trend that has been observed from our users that what kind of a device user
5:12
users that what kind of a device user
5:12
users that what kind of a device user what do they use it on the development
5:15
what do they use it on the development
5:15
what do they use it on the development side nearly 70% of the retail AI
5:19
side nearly 70% of the retail AI
5:19
side nearly 70% of the retail AI research today on deep learning we use
5:22
research today on deep learning we use
5:22
research today on deep learning we use frameworks like bird fork which applies
5:25
frameworks like bird fork which applies
5:25
frameworks like bird fork which applies transform based model to understand user
5:29
transform based model to understand user
5:29
transform based model to understand user user browsing sequence this along can
5:33
user browsing sequence this along can
5:33
user browsing sequence this along can lead into 23 to 30% improvement in
5:36
lead into 23 to 30% improvement in
5:36
lead into 23 to 30% improvement in recommendation accuracy
5:39
recommendation accuracy
5:39
recommendation accuracy uh one of The recent papers that I've
5:42
uh one of The recent papers that I've
5:42
uh one of The recent papers that I've read about is the
5:45
read about is the
5:45
read about is the major beauty retail industry
5:47
major beauty retail industry
5:47
major beauty retail industry that is using bird based model on the
5:50
that is using bird based model on the
5:50
that is using bird based model on the personalized homepage layout based on
5:52
personalized homepage layout based on
5:52
personalized homepage layout based on the recent research and research terms
5:56
the recent research and research terms
5:56
the recent research and research terms and uh pre purchase which they have seen
5:59
and uh pre purchase which they have seen
5:59
and uh pre purchase which they have seen 22% lith on the clickthrough rates right
6:02
22% lith on the clickthrough rates right
6:02
22% lith on the clickthrough rates right for an user it is very for an app it is
6:05
for an user it is very for an app it is
6:05
for an user it is very for an app it is very important that the way the users
6:07
very important that the way the users
6:07
very important that the way the users are clicking the small models on AI uh
6:11
are clicking the small models on AI uh
6:11
are clicking the small models on AI uh has can change the business altogether.
6:16
has can change the business altogether.
6:16
has can change the business altogether. Okay, let's take a moment to pause here.
6:18
Okay, let's take a moment to pause here.
6:18
Okay, let's take a moment to pause here. Um despite of all this innovation, only
6:22
Um despite of all this innovation, only
6:22
Um despite of all this innovation, only one out of 10 AI development in retail
6:25
one out of 10 AI development in retail
6:25
one out of 10 AI development in retail leads to a real business ROI. Uh this is
6:29
leads to a real business ROI. Uh this is
6:29
leads to a real business ROI. Uh this is pretty surprising whenever we hear the
6:30
pretty surprising whenever we hear the
6:30
pretty surprising whenever we hear the number for the first time. But let's
6:32
number for the first time. But let's
6:32
number for the first time. But let's look why the technical challenges go
6:36
look why the technical challenges go
6:36
look why the technical challenges go above beyond the tuning right we often
6:39
above beyond the tuning right we often
6:39
above beyond the tuning right we often see data pipelining that aren't
6:42
see data pipelining that aren't
6:42
see data pipelining that aren't production ready no food feedback loop
6:45
production ready no food feedback loop
6:45
production ready no food feedback loop for model drift or API that were built
6:48
for model drift or API that were built
6:48
for model drift or API that were built with thought clear consumption strategy
6:51
with thought clear consumption strategy
6:51
with thought clear consumption strategy that's when the lot of product and
6:53
that's when the lot of product and
6:53
that's when the lot of product and program managers and business analyst or
6:56
program managers and business analyst or
6:56
program managers and business analyst or a business leader comes into picture to
6:58
a business leader comes into picture to
6:58
a business leader comes into picture to drive this AI really
7:01
drive this AI really
7:02
drive this AI really for for retail company you know train to
7:06
for for retail company you know train to
7:06
for for retail company you know train to create a right recommendation model but
7:09
create a right recommendation model but
7:09
create a right recommendation model but it wasn't just used because of a front-
7:11
it wasn't just used because of a front-
7:11
it wasn't just used because of a front- end app team right it has to be easy
7:14
end app team right it has to be easy
7:14
end app team right it has to be easy with integrations with API as well now
7:17
with integrations with API as well now
7:18
with integrations with API as well now let's look at practical approach to
7:20
let's look at practical approach to
7:20
let's look at practical approach to making this work in production we saw
7:23
making this work in production we saw
7:23
making this work in production we saw that how one out of 10 are uh the AI
7:27
that how one out of 10 are uh the AI
7:27
that how one out of 10 are uh the AI development tools uh were only working
7:30
development tools uh were only working
7:30
development tools uh were only working So looking at the approach
7:33
So looking at the approach
7:33
So looking at the approach generative AR for forecasting think
7:36
generative AR for forecasting think
7:36
generative AR for forecasting think about inventory system retailers deal
7:40
about inventory system retailers deal
7:40
about inventory system retailers deal with 20 to 50 features like historical
7:43
with 20 to 50 features like historical
7:43
with 20 to 50 features like historical sales holidays even weather.
7:46
sales holidays even weather.
7:46
sales holidays even weather. With LSTM and transformers we can
7:49
With LSTM and transformers we can
7:49
With LSTM and transformers we can generate forwardlooking insight to
7:52
generate forwardlooking insight to
7:52
generate forwardlooking insight to predict what SK used to restock in what
7:56
predict what SK used to restock in what
7:56
predict what SK used to restock in what quantity right and when. For example, um
8:00
quantity right and when. For example, um
8:00
quantity right and when. For example, um some of the big industry like Target or
8:02
some of the big industry like Target or
8:02
some of the big industry like Target or Walmart or including Amazon user demands
8:06
Walmart or including Amazon user demands
8:06
Walmart or including Amazon user demands forecasting model that adjust inventory
8:08
forecasting model that adjust inventory
8:08
forecasting model that adjust inventory in a real time because we just can't
8:10
in a real time because we just can't
8:10
in a real time because we just can't have all the stuffs in the inventory on
8:12
have all the stuffs in the inventory on
8:12
have all the stuffs in the inventory on the warehouse and then we don't know
8:14
the warehouse and then we don't know
8:14
the warehouse and then we don't know when the user going to buy it right so
8:17
when the user going to buy it right so
8:17
when the user going to buy it right so it's important for us to forecast major
8:20
it's important for us to forecast major
8:20
it's important for us to forecast major events like back to school or holiday
8:23
events like back to school or holiday
8:23
events like back to school or holiday season or uh black Friday or Christmas
8:26
season or uh black Friday or Christmas
8:26
season or uh black Friday or Christmas events right So it is well in advance
8:31
events right So it is well in advance
8:31
events right So it is well in advance maybe a month prior or two months prior
8:33
maybe a month prior or two months prior
8:33
maybe a month prior or two months prior the the team start working on to see
8:35
the the team start working on to see
8:35
the the team start working on to see that what the forecasting will look for
8:38
that what the forecasting will look for
8:38
that what the forecasting will look for these event uh real time personalization
8:42
these event uh real time personalization
8:42
these event uh real time personalization archiving 75% of accuracy in
8:45
archiving 75% of accuracy in
8:45
archiving 75% of accuracy in personalized means optimizing for speed
8:49
personalized means optimizing for speed
8:49
personalized means optimizing for speed like using Kafka or microservices or
8:52
like using Kafka or microservices or
8:52
like using Kafka or microservices or retail process user click instantly
8:56
retail process user click instantly
8:56
retail process user click instantly recommend next best offer under 100
8:59
recommend next best offer under 100
8:59
recommend next best offer under 100 milliseconds, right? Um, one of the
9:02
milliseconds, right? Um, one of the
9:02
milliseconds, right? Um, one of the fashion uh conversation rate jump eight
9:05
fashion uh conversation rate jump eight
9:05
fashion uh conversation rate jump eight 18% simply by switching formulas which
9:08
18% simply by switching formulas which
9:08
18% simply by switching formulas which I've seen some of the retails like TU or
9:12
I've seen some of the retails like TU or
9:12
I've seen some of the retails like TU or um anywhere that you see a lot of
9:15
um anywhere that you see a lot of
9:15
um anywhere that you see a lot of discounts that they're putting in
9:17
discounts that they're putting in
9:17
discounts that they're putting in because you just click through it and
9:18
because you just click through it and
9:18
because you just click through it and they know that they want you to buy it.
9:20
they know that they want you to buy it.
9:20
they know that they want you to buy it. So they uh make sure that you know put
9:22
So they uh make sure that you know put
9:22
So they uh make sure that you know put with a nice offer. So this is how we do
9:26
with a nice offer. So this is how we do
9:26
with a nice offer. So this is how we do it in the real world and scalable
9:29
it in the real world and scalable
9:29
it in the real world and scalable architectures using Kubernetes. You can
9:32
architectures using Kubernetes. You can
9:32
architectures using Kubernetes. You can horizontally scale workloads during peak
9:35
horizontally scale workloads during peak
9:35
horizontally scale workloads during peak scales events like as I said black
9:37
scales events like as I said black
9:37
scales events like as I said black Friday at
9:40
Friday at
9:40
Friday at we built internal tools that the scale
9:42
we built internal tools that the scale
9:42
we built internal tools that the scale models serving up and down based on the
9:46
models serving up and down based on the
9:46
models serving up and down based on the requested volume to ensure consistency.
9:49
requested volume to ensure consistency.
9:49
requested volume to ensure consistency. Let us see without overpromising. We
9:51
Let us see without overpromising. We
9:51
Let us see without overpromising. We have seen in the last few years before
9:54
have seen in the last few years before
9:54
have seen in the last few years before the pages used to load more when there's
9:57
the pages used to load more when there's
9:57
the pages used to load more when there's an offer we're not able to click or we
10:00
an offer we're not able to click or we
10:00
an offer we're not able to click or we have seen that how do I proceed further
10:02
have seen that how do I proceed further
10:02
have seen that how do I proceed further but currently because uh thanks to IML
10:05
but currently because uh thanks to IML
10:05
but currently because uh thanks to IML model that you can see because of
10:07
model that you can see because of
10:07
model that you can see because of Kubernetes
10:09
Kubernetes
10:09
Kubernetes uh with the current industry that we
10:12
uh with the current industry that we
10:12
uh with the current industry that we have it loads faster and we see the
10:14
have it loads faster and we see the
10:14
have it loads faster and we see the results faster.
10:16
results faster.
10:16
results faster. ML ops pipelines for retail. Let's
10:19
ML ops pipelines for retail. Let's
10:19
ML ops pipelines for retail. Let's really zoom into it to see the backbone
10:23
really zoom into it to see the backbone
10:23
really zoom into it to see the backbone that powers all this. Right? We started
10:26
that powers all this. Right? We started
10:26
that powers all this. Right? We started with data collection and ETL pulling
10:29
with data collection and ETL pulling
10:29
with data collection and ETL pulling customer and product data into feature
10:31
customer and product data into feature
10:31
customer and product data into feature stories.
10:33
stories.
10:33
stories. Right? It's important that what we are
10:35
Right? It's important that what we are
10:35
Right? It's important that what we are building if you're not making into
10:36
building if you're not making into
10:36
building if you're not making into feature just we are building something
10:39
feature just we are building something
10:39
feature just we are building something then if the requirements are not written
10:42
then if the requirements are not written
10:42
then if the requirements are not written well then we cannot proceed. So it's
10:44
well then we cannot proceed. So it's
10:44
well then we cannot proceed. So it's important the data collections and
10:47
important the data collections and
10:47
important the data collections and processing in the ETL CI/CD plannings
10:51
processing in the ETL CI/CD plannings
10:51
processing in the ETL CI/CD plannings allow each model interaction to go
10:54
allow each model interaction to go
10:54
allow each model interaction to go through validations right it's important
10:57
through validations right it's important
10:57
through validations right it's important that we have a pipeline so you're not
10:59
that we have a pipeline so you're not
10:59
that we have a pipeline so you're not doing anything manually being in the uh
11:02
doing anything manually being in the uh
11:02
doing anything manually being in the uh AI world now so using like tensorflow pi
11:05
AI world now so using like tensorflow pi
11:06
AI world now so using like tensorflow pi torch and ab testing it's important that
11:09
torch and ab testing it's important that
11:09
torch and ab testing it's important that how how much the talk is happening with
11:11
how how much the talk is happening with
11:11
how how much the talk is happening with AB testing we make sure that you know
11:14
AB testing we make sure that you know
11:14
AB testing we make sure that you know giving different insight to our users to
11:17
giving different insight to our users to
11:17
giving different insight to our users to see that what they exactly wanted. Uh do
11:20
see that what they exactly wanted. Uh do
11:20
see that what they exactly wanted. Uh do you like the button to be blue, black,
11:23
you like the button to be blue, black,
11:23
you like the button to be blue, black, red, green? So giving all that to the
11:25
red, green? So giving all that to the
11:25
red, green? So giving all that to the different users to see that who is
11:28
different users to see that who is
11:28
different users to see that who is clicking which color is also is driven
11:30
clicking which color is also is driven
11:30
clicking which color is also is driven by AI. That's when the technology is
11:33
by AI. That's when the technology is
11:33
by AI. That's when the technology is going yeah like Docker and Kubernetes
11:36
going yeah like Docker and Kubernetes
11:36
going yeah like Docker and Kubernetes help to push these models into
11:38
help to push these models into
11:38
help to push these models into production at scale. A system like this
11:41
production at scale. A system like this
11:41
production at scale. A system like this monitors retain on drift and keeps the
11:43
monitors retain on drift and keeps the
11:43
monitors retain on drift and keeps the performance on check which is the beauty
11:46
performance on check which is the beauty
11:46
performance on check which is the beauty of the MLOps pipelines.
11:50
of the MLOps pipelines.
11:50
of the MLOps pipelines. Moving on. Now let's talk about the API
11:52
Moving on. Now let's talk about the API
11:52
Moving on. Now let's talk about the API because they they are where your models
11:56
because they they are where your models
11:56
because they they are where your models meet the user, right? You the
11:58
meet the user, right? You the
11:58
meet the user, right? You the interaction between the API um and the
12:01
interaction between the API um and the
12:01
interaction between the API um and the UI which really shows you that what
12:04
UI which really shows you that what
12:04
UI which really shows you that what exactly it is. Great APIs are faster,
12:08
exactly it is. Great APIs are faster,
12:08
exactly it is. Great APIs are faster, secures and flexible. REST is great for
12:11
secures and flexible. REST is great for
12:11
secures and flexible. REST is great for simple product calls. You know, GraphQL
12:14
simple product calls. You know, GraphQL
12:14
simple product calls. You know, GraphQL is a perfect perfect for nested
12:17
is a perfect perfect for nested
12:17
is a perfect perfect for nested structures like outfit
12:20
structures like outfit
12:20
structures like outfit gre ideas for internal high-speed micros
12:22
gre ideas for internal high-speed micros
12:22
gre ideas for internal high-speed micros service communication.
12:24
service communication.
12:24
service communication. But it's important don't not it does not
12:26
But it's important don't not it does not
12:26
But it's important don't not it does not forget catching authentication and rate
12:30
forget catching authentication and rate
12:30
forget catching authentication and rate limiting of all which impact the
12:32
limiting of all which impact the
12:32
limiting of all which impact the customer experience. So keeping the
12:35
customer experience. So keeping the
12:35
customer experience. So keeping the security privacy of a user with that
12:39
security privacy of a user with that
12:39
security privacy of a user with that giving them the faster and more accurate
12:43
giving them the faster and more accurate
12:43
giving them the faster and more accurate system is important. So APIs plays a
12:46
system is important. So APIs plays a
12:46
system is important. So APIs plays a very major role in that. So considering
12:48
very major role in that. So considering
12:48
very major role in that. So considering this is a one of the major major things
12:51
this is a one of the major major things
12:51
this is a one of the major major things for any of the product development
12:55
for any of the product development
12:55
for any of the product development before any model is trained the
12:57
before any model is trained the
12:57
before any model is trained the real-time battle is won or lost the data
13:00
real-time battle is won or lost the data
13:00
real-time battle is won or lost the data prep-processing
13:02
prep-processing
13:02
prep-processing use feature tools to automate feature
13:05
use feature tools to automate feature
13:05
use feature tools to automate feature engineering or DQ validation on the
13:08
engineering or DQ validation on the
13:08
engineering or DQ validation on the large scale attribution sets embedding
13:11
large scale attribution sets embedding
13:11
large scale attribution sets embedding tools
13:13
tools
13:13
tools transforming logs into dense vector that
13:16
transforming logs into dense vector that
13:16
transforming logs into dense vector that boost the like quality to 15 to 20%.
13:21
boost the like quality to 15 to 20%.
13:21
boost the like quality to 15 to 20%. As I've given an example of Kafka, the
13:24
As I've given an example of Kafka, the
13:24
As I've given an example of Kafka, the Kafka streams allows real-time feature
13:27
Kafka streams allows real-time feature
13:27
Kafka streams allows real-time feature updates every interaction, you know,
13:29
updates every interaction, you know,
13:29
updates every interaction, you know, increasing the personalization once you
13:32
increasing the personalization once you
13:32
increasing the personalization once you click on if you go to amazon.com if you
13:35
click on if you go to amazon.com if you
13:35
click on if you go to amazon.com if you click on the pencil, it also give the
13:38
click on the pencil, it also give the
13:38
click on the pencil, it also give the real-time examples of the other pencils
13:40
real-time examples of the other pencils
13:40
real-time examples of the other pencils that you can buy it from, right? or the
13:42
that you can buy it from, right? or the
13:42
that you can buy it from, right? or the fancier ones or which you never thought
13:45
fancier ones or which you never thought
13:45
fancier ones or which you never thought that is there in the market right now.
13:47
that is there in the market right now.
13:47
that is there in the market right now. So that in the back end maybe Kafka or
13:51
So that in the back end maybe Kafka or
13:51
So that in the back end maybe Kafka or active MQ or some of the processing are
13:54
active MQ or some of the processing are
13:54
active MQ or some of the processing are helping to make the system smarter.
13:58
Model strategy for you know real time
14:01
Model strategy for you know real time
14:01
Model strategy for you know real time personalization your development
14:03
personalization your development
14:03
personalization your development strategy must fit your model's purpose.
14:06
strategy must fit your model's purpose.
14:06
strategy must fit your model's purpose. If you're making something and it is not
14:08
If you're making something and it is not
14:08
If you're making something and it is not reaching the purpose and there's no
14:10
reaching the purpose and there's no
14:10
reaching the purpose and there's no point of having that. Uh some of them
14:12
point of having that. Uh some of them
14:12
point of having that. Uh some of them are tensorflow serving or onx real time
14:15
are tensorflow serving or onx real time
14:15
are tensorflow serving or onx real time help with the interface. Docker
14:18
help with the interface. Docker
14:18
help with the interface. Docker containers provide um reproductibility
14:22
containers provide um reproductibility
14:22
containers provide um reproductibility right tensorflow's light on edge devices
14:25
right tensorflow's light on edge devices
14:26
right tensorflow's light on edge devices can reduce latency up to 60% of a
14:28
can reduce latency up to 60% of a
14:28
can reduce latency up to 60% of a realtime in store interaction. And of
14:31
realtime in store interaction. And of
14:31
realtime in store interaction. And of course, AB testing lets you trial
14:33
course, AB testing lets you trial
14:33
course, AB testing lets you trial multiple versions automatically roll
14:36
multiple versions automatically roll
14:36
multiple versions automatically roll back when it is needed. Um, again going
14:39
back when it is needed. Um, again going
14:39
back when it is needed. Um, again going back to the same example that I was
14:41
back to the same example that I was
14:41
back to the same example that I was going on the different colors of the
14:43
going on the different colors of the
14:43
going on the different colors of the buttons, the AB testing would see that
14:46
buttons, the AB testing would see that
14:46
buttons, the AB testing would see that uh if the users are buying if the button
14:48
uh if the users are buying if the button
14:48
uh if the users are buying if the button is more of a green because they feel it
14:50
is more of a green because they feel it
14:50
is more of a green because they feel it is more relatable or the lighter color.
14:53
is more relatable or the lighter color.
14:53
is more relatable or the lighter color. So then we can switch behind for the
14:55
So then we can switch behind for the
14:55
So then we can switch behind for the people that who had a red and giving out
14:56
people that who had a red and giving out
14:56
people that who had a red and giving out the green will they click through it. So
14:59
the green will they click through it. So
14:59
the green will they click through it. So those kind of a testing really helps
15:00
those kind of a testing really helps
15:00
those kind of a testing really helps here.
15:03
here.
15:03
here. Monitoring framework let's not forget
15:06
Monitoring framework let's not forget
15:06
Monitoring framework let's not forget this metrics. Um on this technical side
15:10
this metrics. Um on this technical side
15:10
this metrics. Um on this technical side track frictions recall latency
15:13
track frictions recall latency
15:13
track frictions recall latency throughout drift and system health. On
15:16
throughout drift and system health. On
15:16
throughout drift and system health. On this business side tie your models with
15:18
this business side tie your models with
15:18
this business side tie your models with the key KPIs is what always I say and
15:21
the key KPIs is what always I say and
15:21
the key KPIs is what always I say and important to know what is the
15:22
important to know what is the
15:22
important to know what is the conversation rate are what is the order
15:25
conversation rate are what is the order
15:25
conversation rate are what is the order value revenue and the engagement.
15:28
value revenue and the engagement.
15:28
value revenue and the engagement. As a TPM, I track revenue attribution
15:31
As a TPM, I track revenue attribution
15:31
As a TPM, I track revenue attribution from AB test to ensure value delivery.
15:34
from AB test to ensure value delivery.
15:34
from AB test to ensure value delivery. If we don't have the value for the
15:36
If we don't have the value for the
15:36
If we don't have the value for the delivery, then there's no point of
15:38
delivery, then there's no point of
15:38
delivery, then there's no point of having it, right? Just because we have a
15:42
having it, right? Just because we have a
15:42
having it, right? Just because we have a time and we want to show it to make sure
15:44
time and we want to show it to make sure
15:44
time and we want to show it to make sure that it is automated. But if it is
15:46
that it is automated. But if it is
15:46
that it is automated. But if it is taking more time than the general it is,
15:48
taking more time than the general it is,
15:48
taking more time than the general it is, then there's no point of it. So, it's
15:50
then there's no point of it. So, it's
15:50
then there's no point of it. So, it's very very very important to have uh tie
15:53
very very very important to have uh tie
15:53
very very very important to have uh tie the models with key KPIs.
15:59
Okay. So I put this one on the last to
16:02
Okay. So I put this one on the last to
16:02
Okay. So I put this one on the last to see that how key to successful AI
16:04
see that how key to successful AI
16:04
see that how key to successful AI implementation work. So how do we ensure
16:07
implementation work. So how do we ensure
16:07
implementation work. So how do we ensure long-term success right? That is one of
16:09
long-term success right? That is one of
16:09
long-term success right? That is one of the things that we always ask anything
16:11
the things that we always ask anything
16:11
the things that we always ask anything that we build. First of all very
16:14
that we build. First of all very
16:14
that we build. First of all very important build across functional teams
16:16
important build across functional teams
16:16
important build across functional teams align engineers scientist and analysts
16:19
align engineers scientist and analysts
16:19
align engineers scientist and analysts from the start from the beginning that
16:21
from the start from the beginning that
16:21
from the start from the beginning that you thought you start going to build the
16:24
you thought you start going to build the
16:24
you thought you start going to build the some of the AI tools or internal
16:26
some of the AI tools or internal
16:26
some of the AI tools or internal external or you want to link it to the
16:28
external or you want to link it to the
16:28
external or you want to link it to the APIs anything that we are building get
16:30
APIs anything that we are building get
16:30
APIs anything that we are building get together everybody make sure everyone is
16:33
together everybody make sure everyone is
16:33
together everybody make sure everyone is on the same plate
16:35
on the same plate
16:35
on the same plate and use CI/CD and AB testing to release
16:40
and use CI/CD and AB testing to release
16:40
and use CI/CD and AB testing to release models faster. So we're not taking we're
16:42
models faster. So we're not taking we're
16:42
models faster. So we're not taking we're in the agile world. Uh it is not like a
16:44
in the agile world. Uh it is not like a
16:44
in the agile world. Uh it is not like a waterfall that we used to do before. So
16:47
waterfall that we used to do before. So
16:47
waterfall that we used to do before. So make sure that you're going fast and
16:49
make sure that you're going fast and
16:49
make sure that you're going fast and depending on the AB testing trial with
16:51
depending on the AB testing trial with
16:51
depending on the AB testing trial with users. Do not give the full-fledged
16:53
users. Do not give the full-fledged
16:53
users. Do not give the full-fledged which we are not sure that whether our
16:56
which we are not sure that whether our
16:56
which we are not sure that whether our users will like it or not. Customize
16:58
users will like it or not. Customize
16:58
users will like it or not. Customize everything and prioritize data quality
17:02
everything and prioritize data quality
17:02
everything and prioritize data quality is governance, validation and feature
17:04
is governance, validation and feature
17:04
is governance, validation and feature stories. Make sure the data is right.
17:07
stories. Make sure the data is right.
17:08
stories. Make sure the data is right. feed more data to the system as much as
17:10
feed more data to the system as much as
17:10
feed more data to the system as much as you can. Run on internal uh data for the
17:14
you can. Run on internal uh data for the
17:14
you can. Run on internal uh data for the for the company that we have uh or any
17:17
for the company that we have uh or any
17:17
for the company that we have uh or any data which is available for you. Feed
17:19
data which is available for you. Feed
17:19
data which is available for you. Feed it, try it, validate it really well and
17:23
it, try it, validate it really well and
17:23
it, try it, validate it really well and make sure the feature stories are
17:24
make sure the feature stories are
17:24
make sure the feature stories are aligned with ROI.
17:26
aligned with ROI.
17:26
aligned with ROI. Set a success metrics. Track both
17:29
Set a success metrics. Track both
17:29
Set a success metrics. Track both technical and business performance. Make
17:31
technical and business performance. Make
17:31
technical and business performance. Make sure that you have everything needed to
17:34
sure that you have everything needed to
17:34
sure that you have everything needed to call it as whether this project
17:35
call it as whether this project
17:35
call it as whether this project successful in the EB testing. If it is
17:38
successful in the EB testing. If it is
17:38
successful in the EB testing. If it is not, it's okay to move on then getting
17:40
not, it's okay to move on then getting
17:40
not, it's okay to move on then getting it stacks.
17:42
it stacks.
17:42
it stacks. Okay, this closes the loop between dev
17:44
Okay, this closes the loop between dev
17:44
Okay, this closes the loop between dev work and ROI that how the world has
17:47
work and ROI that how the world has
17:47
work and ROI that how the world has drastically changed for us.
17:50
drastically changed for us.
17:50
drastically changed for us. Yes, to finally to wrap up, AI is not
17:53
Yes, to finally to wrap up, AI is not
17:53
Yes, to finally to wrap up, AI is not just about building smart models. It's
17:55
just about building smart models. It's
17:55
just about building smart models. It's about building systems ones that scale,
17:59
about building systems ones that scale,
17:59
about building systems ones that scale, adopt and deliver real business value.
18:02
adopt and deliver real business value.
18:02
adopt and deliver real business value. The smart divi dev assistant in VS codes
18:05
The smart divi dev assistant in VS codes
18:05
The smart divi dev assistant in VS codes bring development, testing, deployment
18:08
bring development, testing, deployment
18:08
bring development, testing, deployment and monitoring into the intelligent
18:11
and monitoring into the intelligent
18:11
and monitoring into the intelligent workflow which is very important. It's
18:13
workflow which is very important. It's
18:13
workflow which is very important. It's not that what we build, what we maintain
18:16
not that what we build, what we maintain
18:16
not that what we build, what we maintain it, what we monitor it.
18:18
it, what we monitor it.
18:18
it, what we monitor it. I think that's all I had. Thank you so
18:20
I think that's all I had. Thank you so
18:20
I think that's all I had. Thank you so much for your
Retail Trade
Retail Equipment & Technology