Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:00
Each phase should have a deliverable.
0:02
It might not be a specific
0:04
product or anything, is okay to be
0:06
imperfect, but let's have a goal. And
0:09
when we have a goal and when we have a plan
0:12
at each phase, then
0:14
big foundational work
0:16
might seem achievable. although
0:18
you cannot see an ROI at 1
0:20
and 2, but since I know I
0:23
can give you that ROI back, at
0:25
phase six, so that this is very important.
0:27
In that case, you'll be more convinced than
0:30
me going to you and saying that, Hey, I don't
0:32
really know how to go to phase six, but we need to
0:34
do phase one.
0:36
How can you drive developer productivity,
0:38
lower costs and deliver better
0:40
products? On June 20th and
0:42
27th LinearB is hosting a workshop
0:45
that explores the software engineering intelligence
0:47
category. We'll showcase how data-driven
0:49
insights and innovative workflow automations
0:51
can optimize your software delivery practices.
0:54
At the end of the workshop, you'll leave with a complimentary
0:57
Gartner market guide to software
0:59
engineering, intelligence, and other resources to
1:01
help you get started. Head to the show notes
1:03
or LinearB dot IO slash events
1:05
to sign up today.
1:06
Welcome back to Dev Interrupted, I'm your co host Conor
1:09
Bronsdon. Today I'm being joined by Sanghamitra Goswami,
1:12
Senior Director of Data Science and Machine Learning
1:14
at PagerDuty. Sanghamitra,
1:16
thank you so much for joining me.
1:18
Thank you, Conor, for inviting me.
1:20
Honestly, it's my pleasure. I think we
1:22
could really benefit from your expertise as
1:24
someone who has such a deep understanding
1:27
of the approach that data scientists have taken
1:29
to developing AI models. And,
1:31
I mean, frankly, AI is all the rage right now, right?
1:33
Everyone's talking about it. Everyone has opinions
1:36
on what it is. And so it's important that
1:38
we level set with the audience a bit, and
1:40
have the opportunity here to pick the brain of a
1:43
data science leader like you, and understand
1:45
how engineering teams can
1:47
translate this and leverage
1:49
AI models, or LLMs in
1:52
particular, within their org. Let's
1:55
maybe unravel some of the strategies
1:57
that champion the role of data science,
1:59
machine learning, and AI teams, and help our
2:01
audience understand how to navigate this
2:04
emerging future and ever expanding landscape.
2:06
Why don't we talk a bit about the history of LLMs,
2:09
what they are. and why they're so important.
2:12
You know, Conor, it's been a crazy time
2:14
now. Um, nine months back
2:16
when ChildGBD came out, uh,
2:18
PagerDuty leadership, they told me, Mitra, we need
2:20
to do something with, uh, LLMs, and it,
2:23
it's crazy the way the world is saying
2:25
that, hey, we, we want to do LLM, let's
2:27
have a feature that uses
2:29
LLMs. So what are LLMs?
2:32
Large language models, they leverage
2:34
foundational machine learning, AI, deep
2:36
learning. Models to understand
2:38
natural language and to give answers
2:41
as so they can talk to us like a bot
2:44
You look at the history of NLP, it's based on
2:46
NLP, it's based on the NLP models
2:48
that we have built out. If you look at the history
2:50
of NLP, in 1948, Shannon
2:52
and Weaver, the first paper came out.
2:55
And, and, during that time,
2:57
we didn't have the computer storage that is
2:59
possible now. We didn't, we couldn't
3:01
really have a lot of computational
3:04
power. So, It
3:06
was not possible to always run
3:09
these large language models because they require
3:11
large amounts of
3:13
text. Right. However,
3:15
from that start, where Shannon
3:18
and we were in 1948, if you fast forward,
3:21
um, I would actually mention
3:23
there is another milestone
3:25
where the transformer architecture got
3:27
introduced and, um,
3:30
attention is all you need is the paper. So,
3:32
if I look at these two milestones and how
3:34
the landscape has changed, the
3:36
computational power, GPUs.
3:39
So with everything in mind, now is the
3:41
perfect time that we can all
3:43
reap the benefits of years
3:45
of research and computational
3:48
power. And engineering in that
3:50
sense. So this is the perfect time.
3:52
Absolutely. It's, it's really interesting to think about
3:54
how, even just a few years ago before we realized
3:57
that we could paralyze processes within
3:59
AI model development, uh, using
4:01
GPUs, there just wasn't
4:03
this speed of development that we've seen around AI models
4:05
today. I know. Uh, so I, I'd
4:08
love to kind of talk about how
4:11
teams can actionably leverage AI
4:13
or LLMs, uh, within their tooling. What
4:15
would be your advice to engineering leaders who
4:18
are thinking, Hey, I wanna start using
4:20
AI to extend the capacity of our product.
4:22
How should they kick off?
4:24
I think, let's start with ai. Just
4:26
ai. Someone wants to do AI
4:28
with their teams. Okay. That is a
4:30
huge pro, huge challenge because
4:33
if you look at all the leaders
4:35
in the industry, how do we
4:37
see if something is successful in the industry?
4:39
We have to get some ROI with our,
4:42
with all the endeavors, right? And
4:45
data science, AI, it's an experiment.
4:48
So when we start building it, it is always
4:51
not clear how this is going to show up.
4:54
For example, in AI models, we,
4:56
we try to build a model. We experimented
4:59
with historical data, and then we take
5:01
the model out in the market. And
5:03
then it starts getting all this new data from
5:05
our customers and our
5:07
model is giving answers
5:10
live in the run time. So
5:12
it's very difficult. It's an experiment that we are
5:14
running. Yeah. Leaders
5:18
should be very careful that they know
5:20
that what are the risks of running these
5:22
experiments. That's number one. Number
5:25
two, they should be very careful on
5:27
how they measure adoption
5:29
or how the results of these experiments
5:32
are being used by end users. Once
5:35
these two pillars are in place,
5:37
I think it's easier for leaders to measure
5:39
value and define ROI. Without
5:42
these two, AI, we can't do AI
5:44
or we can't do So
5:48
these are the two main, I would say, pillars
5:50
if you want to do AI. Now there are
5:52
other, uh, you know, organizational
5:55
challenges as well. For example,
5:58
democratization of data. Everybody
6:01
talks about it. But it's
6:03
a difficult job. Their data is
6:05
always in silo. There are different channels.
6:08
If you look at marketing, you get,
6:10
there is social media, there are emails,
6:13
there is YouTube or other
6:15
social networks. So, so many different
6:17
channels and you have to get the data together.
6:20
If you look at logistics, there
6:22
is carrier data in the ocean, carrier
6:24
data in the Um, and on
6:26
road, in, in air.
6:29
So it's just that if we
6:31
want to measure any process,
6:33
data is always in silo and
6:35
there is a huge, there are huge effort that
6:37
need, that is needed, um, on
6:39
the side of the person
6:41
who wants to do AI, to do a successful
6:44
AI. That's number one.
6:46
And we need, we have the architecture,
6:48
we have the solution, we know
6:50
that we need a data lake or we need
6:52
a source of data where all the data can
6:54
be consolidated. But it's, it's
6:57
a huge effort. It's a huge
6:59
foundational effort that always
7:01
doesn't show direct ROI. So
7:04
you have to convince as a leader in AI,
7:07
you have to convince your leaders that, hey,
7:09
I'm going to do this foundational effort. However,
7:12
you might not see a direct Right
7:15
now, but it's going to,
7:18
um, you know, executives talk about this
7:20
flywheel effect, but it's going to create
7:22
that flywheel effect at some point. So.
7:26
That's the discussion they need to drive a
7:28
Absolutely. I think it's really important for us to
7:30
understand both the potential and the risks
7:32
of ai. And, and I don't mean,
7:35
you know, the risks of a GI
7:37
and you know, this world
7:39
of like, oh, like crazy
7:41
cyber villain and creative ai like that. That's, that's
7:44
fun to think about Unci. Yeah, sure, we can talk about that.
7:46
But like, like more specifically to your business. There
7:49
is a challenge where if you don't put these foundations
7:51
in place, there's major risk to how
7:53
your business will present itself, whether that model will hallucinate.
7:55
And this is where it comes down to these foundational
7:58
data science concepts you're talking about, like, is
8:00
my data siloed? Do we have the
8:02
right training data? Is that training data
8:04
validated? Are there issues with
8:06
that data set that are going to cause long term
8:09
issues? And when I've talked
8:11
to other data science leaders like
8:13
yourself, that is one of the things people really hone
8:15
in on. So I'm glad you bring up this foundational piece
8:18
because A lot of leaders are getting
8:20
pushed by their board or pushed by their c-suite
8:22
of like, oh, we need to get AI in their product. But if,
8:24
if the data that you're feeding in to train
8:26
the model isn't, some data, isn't
8:29
data that is actually, uh, validated
8:31
and maybe peer reviewed or, or, uh,
8:33
checked on, like there are major risks that you put in play.
8:35
Yes. Yes, absolutely. You need
8:37
to know what your data can do. Without
8:39
that, you know, garbage in, garbage out.
8:42
You cannot save yourself even with an
8:44
LLM. So, yeah.
8:46
Very well said, and I think
8:48
it's challenging though for a lot of leaders
8:51
to get buy in on that foundational work
8:53
as you point out because there is an immediate ROI.
8:55
So, how should
8:57
engineering leaders start to, uh,
9:00
get that buy in about
9:02
ensuring they actually do all the steps needed to
9:04
be successful? Yes,
9:06
I think, we work in an agile
9:08
world, so we should have
9:11
a plan. We should have a plan with different
9:13
phases. And I always
9:15
say this to my team that each phase should
9:17
have a deliverable. It might not
9:19
be a specific product
9:22
or anything, but it should have a goal. It should
9:24
have a deliverable. I don't know
9:26
if I read about this Wabi
9:28
Sabi. It's a Japanese, uh,
9:31
Guiding Principle. It talks about continuous
9:34
development and that imperfection
9:36
is good. And I say this to my
9:39
team that while we say let's do something,
9:41
it is okay to be imperfect,
9:44
but let's have a goal. And when
9:46
we have a goal and when we have a plan
9:48
at each phase, then
9:50
big foundational work
9:52
might seem achievable. And that
9:54
is very important as Say
9:57
you are my boss, Conor, and I'm talking
9:59
to you, and I'm giving you a plan, and I'm saying
10:01
that, hey, phase 1, 2, I have
10:03
a goal, and I know I can go to phase
10:05
6, and at the end of, although
10:07
you cannot see an ROI at 1
10:10
and 2, but since I know I
10:12
can give you that ROI back, at
10:14
phase six, so that this is very important.
10:17
If I give you the plan that in that
10:19
case, you'll be more convinced than me
10:21
going to you and saying that, Hey, I don't really
10:23
know how to go to phase six, but we need to do
10:25
phase one.
10:26
So if I'm an engineering leader who hasn't
10:28
had deep experience with data science or ai,
10:31
and I'm thinking about how do I build this phased approach. Um,
10:34
what would be the general
10:36
steps you would advise, or is there a resource where
10:38
leaders can go in and say, Hey, let me look at,
10:40
I don't know, a template to start applying
10:42
to our specific use case?
10:45
Yes, I think there are many if you, if you look
10:47
at Google, like, how do you use, use gather
10:50
ROI for data science projects? But
10:52
I would say before we do that, it is very
10:55
important that in any organization, there is
10:57
a product counterpart with a data science
10:59
engineering manager. I believe
11:01
there should be other people championing
11:04
data science rather rather than
11:06
get So you in Yeah. To get buy-in rather than the data
11:08
scientists themselves. So it is critical
11:10
that you have a friend in the product organization
11:13
because they can look at the product holistically
11:16
from a top level view and they
11:18
can help you. Go ahead.
11:20
What would you say to people who are having trouble
11:22
finding that champion or picking the right champion?
11:26
Convince your boss, convince your boss
11:28
that you need a product partner. A single
11:30
person or a group of data scientists
11:32
can't always do everything. You need
11:34
to have a, someone else beyond that organization
11:37
who could champion for you. Sometimes
11:39
an executive can play that role too,
11:41
but with, with
11:43
having some time and common goals
11:46
and everything, I think it's critical that data
11:48
science organizations have a product partner.
11:50
And I'm sure this creates some translation
11:53
challenges across the company, as you're
11:55
trying to bring in these other stakeholders and get people
11:57
to buy in because you know you need the support,
11:59
but maybe the goals across
12:02
those different organizations can be different. Yes. Um,
12:04
how would you try to solve that
12:07
cross organizational translation challenge
12:09
to, to get these champions?
12:11
Well, I, I don't think the data scientists
12:13
can solve it by themselves. What, what
12:15
they can do is they, they
12:17
can say that, hey, I understand
12:19
it's late in your roadmap and I can
12:22
be ready on my part and I can make
12:24
it easier for you to understand
12:27
and access what I'm developing.
12:29
But I do think executives play a very
12:32
important role here because they need
12:34
to drive alignment across different
12:36
teams at the organization. Let's say
12:38
engineering team A and engineering team
12:40
B both wants to do data science, but
12:43
they don't have time because their roadmaps
12:45
are full of other projects. Data,
12:47
whatever the data scientists do, it won't
12:49
convince them, right? So the executives need
12:51
to prioritize that, hey, or, or a product
12:54
partner who can look at it and say, hey, this
12:56
data science feature, if we do it, this
12:58
will drive huge ROI than
13:01
a, than a small change or than
13:03
any other feature that we are taking out
13:05
this year. So we need, we need
13:07
to have executive alignment on roadmap
13:10
across teams and also. Some
13:12
other champions, but what the data scientists
13:15
or data science, uh, data science,
13:17
uh, organization leaders can do is
13:19
they can think of, okay, here
13:21
are some benefits of, um,
13:24
empowering data scientists and
13:26
data engineers so that they can write
13:28
code well. Um, this
13:30
is, I'm going off a tangent because,
13:32
you know, data scientists come from different backgrounds
13:35
and they are always not the best software engineers.
13:38
So they need support from data engineers and
13:40
they need to productionize their code, write production
13:42
level codes. So what the leader
13:44
in the data science organization do is make
13:47
sure that the organization
13:49
is empowered to Build
13:51
something that is very easily accessible
13:54
and can be taken by the
13:56
engineering team, and the engineering team doesn't
13:59
spend a lot of time building that or
14:02
understanding that.
14:03
It's interesting because you're talking a lot about these
14:06
change management concepts, frankly, of like,
14:08
you know, getting organizational alignment, building
14:10
up champions within the org, making sure you get
14:12
buy in, so you can showcase that ROI,
14:15
ensuring you have these phased rollouts and a
14:17
clear goal for each step of your process.
14:20
What if you're having trouble
14:22
getting that kind of buy in? Are there
14:24
ways that, you know,
14:26
data science or engineering teams can
14:29
leverage currently available
14:31
AI models or tooling to showcase the
14:33
ROI and then create that buy in?
14:36
Yes, there are also many
14:38
tools in the market that ROI,
14:42
but it's a little bit difficult. Once again, I
14:45
think, uh, It's an
14:47
experiment. That's how I see data science.
14:49
So it's not, depending on how much
14:51
customer, how much data your customer
14:53
have. You know, some customers
14:55
might be new. Some customers might not
14:58
be storing data very well. So
15:00
the experiment that I have run on
15:02
my end might not always
15:05
be great when I run it
15:07
with real time data with all my
15:09
customers. So having those risk
15:11
factors. Figuring those
15:13
out before release or having
15:15
a slow release so that you can talk to your
15:17
customers and figure out. Conor is
15:19
a great customer because we have five
15:22
years of data and our model is going to
15:24
give very good results. Whereas Mitra
15:26
might not be a good customer because she
15:28
has only six months of data. Figuring
15:31
those out and how, what is the fraction of your
15:33
customers will give good results. So the,
15:35
the risks, I think thinking those
15:38
ahead of time. Makes a huge
15:40
difference. I mean,
15:41
we've talked about this some on the show before about the importance
15:43
for leaders of understanding the risks of
15:45
even exciting opportunities. Yes, and I
15:48
think you bring up a good one, which is like, it's really easy
15:50
for us to over-exaggerate the
15:52
impact of AI on a particular customer
15:55
or on, you know, a feature or product
15:57
where maybe the realistic. Truth is that it's
15:59
going to take time for it to develop because you need that,
16:02
uh, data integrity to, to actually build
16:04
up, because you need more customer data. How
16:06
should you go about communicating with customers
16:08
about what you're able to do with AI
16:10
as you build your program?
16:12
I think building trust with customers is
16:14
key. Um, at
16:17
PagerDuty we have a process
16:19
called Early Access, where
16:22
our product is not fully built out, but
16:24
we are in the Early Access program, we have a prototype,
16:27
and we can Ask our customers to
16:29
use it and give us feedback. I think that
16:31
feedback is critical. They can tell
16:33
us that, hey, it's giving great results. They
16:35
can tell us it's giving very bad results. So
16:37
then we know and we can improve
16:40
and we can, so this early access
16:42
program is very useful.
16:44
How are you leveraging AI at PagerDuty?
16:47
We do a lot of AI. Um,
16:49
so We have five
16:52
features, and when I say features, these are
16:54
features which have different models in the, um,
16:56
in the backend. So we have five AIOps
16:59
features, and our AIOps, which used
17:01
to be an add on for our
17:03
full IR product, now is a separate BU.
17:06
So we have, um, A lot of
17:08
features in AIOps, noise reduction. I was
17:10
just talking to someone who mentioned that, uh,
17:13
uh, it's always a problem when you have
17:15
lots of, um, alerts.
17:18
And we are talking about security camera, like Google
17:20
camera. You keep on getting alerts and then
17:22
you are lost, right? So, the same thing
17:24
happens when people use PagerDuty. People
17:26
use PagerDuty when there is an incident and you are
17:28
getting alerts. And if you get a lot of
17:30
alerts, if you're inundated by alerts, you don't
17:32
know which one to go for. So we
17:34
have very good noise reduction
17:37
algorithms and we use AI
17:40
to, um, build those,
17:42
uh, noise reduction algorithm. That's, that's
17:44
a super smart use case. 'cause that kind of
17:46
cognitive load, it really makes us start to tune out.
17:48
I mean, like, I'm sure we've all been guilty of this.
17:51
Maybe with our email. Sometimes it's, ah,
17:53
so many, okay, I just gotta get through these things. Yes. And
17:56
it's so easy to miss something that might be important
17:58
if you're not really staying on it. Uh,
18:00
that's a great example of how you can leverage the power
18:02
of AI to assist for your customers.
18:05
Um, are there other ways that you see PagerDuty
18:08
leveraging AI in the future? Yes, we
18:10
have root cause. So these are, I'm talking
18:12
about, you know, non NLM AI
18:14
ops features. We have root cause,
18:17
we have probable origin. Uh,
18:20
what we do with these features is during
18:22
the incident, during the triage process, we try
18:24
to provide information to developers who are
18:26
looking We're trying to figure out what has
18:29
gone wrong, how can we figure
18:31
out, how can we resolve the incident faster.
18:33
So we have a suite of features
18:35
on that end. On the LLM side
18:37
of things, we have three new features
18:39
that are coming out. These are our first
18:42
Gen AI features.
18:44
We have a summarization use case.
18:47
I think this is a very good use case and
18:49
uh, one of the ways I always say,
18:51
once again going off a little tangent,
18:53
I always say that if you want to do LLM, find
18:55
a good use case. And I think this is
18:57
an awesome use case. So During
19:00
that incident, developers are trying
19:02
to solve, you know, a problem
19:04
and say that, okay, I'm resolving the incident. But
19:06
even during that phase, they have to update
19:09
their stakeholders or external,
19:11
uh, companies that, who are waiting for
19:14
information about the incident that is going on. That's
19:17
a very difficult job. Developers
19:20
who are always in the back end, they need to
19:22
write up an email and, you know, divide
19:24
their attention between solving a problem
19:27
and drafting up an email. So you'll give it
19:29
to the generative AI platform because
19:31
now they can do it for yourself. And
19:34
those conversations are already
19:36
there in Slack, in Zoom, in Microsoft
19:38
Teams. So why repeat it? Ask
19:40
your generative AI model to write it for yourself.
19:43
Smart.
19:43
So I think this is a very, very good use
19:45
case. and empowering to,
19:48
to the developers who are using PagerDuty.
19:51
And you mentioned this idea of ensuring
19:53
you find the right use cases or good use cases.
19:55
What's the approach that
19:57
you should think people should take to that?
19:59
What is the problem that you are solving for?
20:02
How would you provide relief to your
20:04
customers or
20:06
the end users? What would they find
20:08
useful? That's the key. And that's true
20:10
for LLMs too. You want to find
20:12
the use case, but is your use case
20:14
solving the problem that is being asked
20:17
by a lot of your customers?
20:19
And that's just good business advice, period, right?
20:21
Like, solve your customer problems. The
20:23
phrase, like, make your beer taste better has
20:26
been popularized recently. It's a great example. We have,
20:28
we say in PagerDuty,
20:30
champion the customers.
20:32
Great way to put it. I'd love to just get some more general
20:34
thoughts from you about your viewpoint
20:36
on AI in general,
20:38
where the industry's going, the explosion of success
20:41
here now that we have paralyzed GPU
20:43
models, we have years of them working, um,
20:46
obviously there's been an explosion in the public consciousness
20:48
of the ability to leverage AI,
20:50
and as you pointed out, like, this
20:53
all goes back to Like,
20:55
old papers, these, this goes back to
20:58
old sci fi novels, frankly, where we talked about
21:00
these ideas. Yes. And, and now we're
21:02
seeing them come into reality. Uh,
21:04
what are some of the things that you're excited about by
21:07
the current AI revolution that's happening?
21:10
I think I'm seeing very good use
21:12
cases. One use
21:14
case that I really loved. I am
21:17
big on Instagram. You know, I
21:19
was looking at the photo editing capabilities
21:22
and you can just take out a person
21:24
that you didn't like.
21:25
So if you've got an ex
21:27
boyfriend or girlfriend or something like
21:30
that, like, yeah.
21:31
No, think about it. I want my
21:33
picture before the Eiffel Tower.
21:36
And I don't want anyone else. And
21:38
I can do that now with AI. So
21:40
I love it. I love some
21:42
of the applications that are coming up. This
21:45
is a fun one, but there are very useful
21:47
ones in the, if I look around. Um,
21:50
recently when I was going to
21:52
the Chicago airport, they
21:54
did a, they did a facial
21:57
scanning. And they didn't actually
21:59
scan my boarding pass. Not LLM,
22:01
but still it's so cool. Where
22:03
I'm just walking and there is a machine
22:05
who is scanning my face.
22:07
I get some privacy concerns that I have to
22:09
admit, but it is very cool.
22:10
Yeah. Yeah, it's just so cool. Yeah.
22:12
Uh, well, thank you so much for taking the
22:14
time to chat with me today about this. It's been fascinating
22:16
to dive into your thoughts about AI. Do you
22:18
have any closing thoughts you'd like to share with the audience
22:21
about either how they should approach LLMs
22:23
or what all of this change means?
22:25
I would say data science leaders
22:27
fight for, fight
22:29
for space. Uh, you need to do more.
22:33
Uh, think of a good use case. Ask
22:35
your executives for a product partner. Try
22:38
to prove that this, the features you want to
22:40
develop, that the use case you are vouching
22:42
for, you want to build for, is going to solve
22:46
a customer problem. And that
22:48
is needed. Write up. I think
22:50
writing is very useful. And
22:52
give it away for people to consider,
22:55
take feedback. Be vocal. Yeah,
22:57
I would say that
22:58
Well said song. thank you so much for coming
23:00
on the show. It's been a distinct pleasure. If
23:03
you're someone who's listening to this conversation, uh,
23:05
consider checking out on YouTube. We're here in the midst
23:07
of an incredible lead dev conference, here in
23:09
Oakland, and I think it would be a ton of fun for
23:11
you to see us having this conversation live, uh,
23:14
on, on the YouTube channel. So that's,
23:16
uh, dev interrupted on YouTube, check it out. And,
23:18
uh, once again, thanks for coming on the show.
23:19
Thank you, Connor. Thanks for the invitation.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More