201: What I ACTUALLY Do as a Data Analyst

Speaker: 00:00:00

Avery Smith-1: I'm a senior data analyst

with 10 plus years of experience.

2

: 00:00:03

What did I do in those 10 years?

3

: 00:00:05

What tools did I use?

4

: 00:00:06

What problems did I solve?

5

: 00:00:08

That is the topic of today's episode,

and I'm gonna tell you everything

6

: 00:00:11

so that way you know what to expect

as a data analyst in the future.

7

: 00:00:15

I've had a really vast career where I've

worked for one of the biggest oil and

8

: 00:00:19

gas companies in the world, and I've

also worked for a 10 person biotech

9

: 00:00:23

startup that you've never heard of.

10

: 00:00:25

Before, so let's get into it.

11

: 00:00:27

By the way, if you're new here, my

name is Avery Smith and I try to

12

: 00:00:29

share useful data content that will

help you start your data career.

13

: 00:00:33

If that's of interest to you, you

gotta check out my newsletter.

14

: 00:00:36

30,000 other aspiring data

analysts are already subscribed.

15

: 00:00:39

Go to data career jumpstart.com/newsletter

16

: 00:00:42

or find the link in the

show notes down below.

17

: 00:00:44

So the first company I wanna

talk about is ExxonMobil.

18

: 00:00:47

And what was it like being a data analyst

and a data scientist at ExxonMobil?

19

: 00:00:51

Obviously this is one of the

biggest companies in the world.

20

: 00:00:53

There's like 70,000 employees and

they do a lot of different things.

21

: 00:00:58

Now, I worked in the downstream.

22

: 00:01:00

Part of the business, which

basically means the refiners.

23

: 00:01:02

These are the people that are taking oil

and turning it into gasoline essentially.

24

: 00:01:06

And what do we do there as data analysts?

25

: 00:01:08

Well, we tried to make a mathematical

model of every single part of the

26

: 00:01:13

refinery, and I don't think this is,

you know, groundbreaking to those who

27

: 00:01:17

are in the oil and gas business or

any sort of manufacturing business.

28

: 00:01:20

If you can create what's called

like a digital twin or like a math

29

: 00:01:23

twin of your process, you'll be able

to experiment with the math model

30

: 00:01:27

instead of experimenting in real life.

31

: 00:01:29

So you can be like, well, if I twisted

this temperature, or I changed this

32

: 00:01:32

pressure, or we, you know, added

this new oil, what would change?

33

: 00:01:35

Would we make more money?

34

: 00:01:36

Would we make less money?

35

: 00:01:37

What would go well?

36

: 00:01:38

What would go poorly instead of actually

experimenting In real life, you can

37

: 00:01:41

experiment with these simulations with

your data model, and that way you don't

38

: 00:01:45

actually have to do it in real life.

39

: 00:01:47

Now to create these models, there's lots

of different ways that you can do them.

40

: 00:01:50

I'm not getting into the

nitty gritty of like.

41

: 00:01:52

Modeling these types of things.

42

: 00:01:54

But when you think model, the simplest

version that you can think of in

43

: 00:01:57

your head is linear aggression.

44

: 00:01:59

And if you're not familiar

with linear aggression, you

45

: 00:02:01

learned it definitely in school.

46

: 00:02:02

It's the simple thing

of Y equals MX plus B.

47

: 00:02:06

That's the simplest form.

48

: 00:02:07

So basically you have an input.

49

: 00:02:09

An X.

50

: 00:02:10

If based upon your input, can you

predict what the output is going to be?

51

: 00:02:14

If it, you know is a linear relationship,

you'll be able to have the slope that's

52

: 00:02:18

the m and some sort of a y intercept,

and basically guess what the output

53

: 00:02:22

the Y is going to be based on the X.

54

: 00:02:24

Now you can do that a

lot more complicated.

55

: 00:02:26

You could do multivariate, linear

regression, which is like y equals.

56

: 00:02:31

M1 X one plus M two X two plus X 3M three.

57

: 00:02:37

Oh, it's so confusing.

58

: 00:02:38

But my whole point here is like we

were doing these mathematical models,

59

: 00:02:41

and the simplest form that you

can think of is linear aggression.

60

: 00:02:43

So I created a lot of these

models as a data analyst.

61

: 00:02:46

And I also used data analytics to try to

understand our simulation results better.

62

: 00:02:53

So we'd actually run dozens,

hundreds, thousands of simulations

63

: 00:02:58

trying, you know, different things.

64

: 00:02:59

Well, what if this pressure went up by a

little bit, or this temperature went down?

65

: 00:03:03

To actually look at a thousand

different results is really hard to do.

66

: 00:03:06

So we used data analytics

to try to understand the

67

: 00:03:08

results a little bit better.

68

: 00:03:09

And a lot of this was done in a

Power BI dashboard, so I used a lot

69

: 00:03:12

of Power BI dashboards right there.

70

: 00:03:14

And to do the modeling.

71

: 00:03:15

We actually did a lot in Excel, believe

it or not, and we did a lot in Python

72

: 00:03:18

and we even used a more proprietary

software that you don't hear a whole lot.

73

: 00:03:22

It's from sas.

74

: 00:03:23

It's called Jump, JNP, to do our modeling.

75

: 00:03:25

So those are the tools that we're using

at Axon, and that's the problem that

76

: 00:03:28

we're trying to solve is basically,

hey, if we wanna make changes inside of

77

: 00:03:31

our huge manufacturing system, can we

actually come up with a way to test it

78

: 00:03:36

before testing it in real life so we can

kind of know and expect what to happen?

79

: 00:03:40

I think that's common for,

you know, manufacturing.

80

: 00:03:43

I think that's common for any sort of

like time series data you might have

81

: 00:03:47

is if you can create a model, it's

useful for the company to be able

82

: 00:03:50

to predict the future and be able to

figure out what's going to happen.

83

: 00:03:53

A lot of the times this type of

analytics is called prescriptive

84

: 00:03:56

analytics, where you're actually like

trying to not predict what's going

85

: 00:03:59

to happen in the future, but trying

to decide if you make these changes.

86

: 00:04:03

How will the system basically be affected?

87

: 00:04:05

The next data job I wanna talk about was

when I was a data analyst at this nano

88

: 00:04:09

biotech startup, like think 10 people.

89

: 00:04:12

When I joined the company, this

company made really cool nano sensors.

90

: 00:04:16

So think of it as almost like a game

boy, uh, game, like from the olden days,

91

: 00:04:21

that's like the size of this little board.

92

: 00:04:22

And on this board there was a bunch

of different sensors this, you

93

: 00:04:26

know, chemistry company had built.

94

: 00:04:28

And the sensors would basically react to

what was in the air and we would track.

95

: 00:04:32

How their electricity basically,

or their, their amperage or their

96

: 00:04:36

current, through these different

sensors would change when these

97

: 00:04:39

different chemicals in the air hit it.

98

: 00:04:41

So, for example, if you were holding

it in the air, you know, all the

99

: 00:04:44

lines would be kind of stagnant.

100

: 00:04:46

But for example, let's say you

brought an orange next to it, it

101

: 00:04:49

would basically smell the orange.

102

: 00:04:51

And each sensor would react differently

to that orange being nearby.

103

: 00:04:55

And when you have, uh, an array of

these 12 different sensors, you can

104

: 00:04:59

basically create the equivalent of

like a fingerprint, but for smells.

105

: 00:05:04

So think of it as like the smelling device

that would basically take smell prints.

106

: 00:05:08

My job as a data analyst there was to

actually look at the time series data.

107

: 00:05:12

'cause we'd run these experiments where

you'd have like basically background

108

: 00:05:15

noise for a certain amount of time

and then you'd introduce something

109

: 00:05:19

like an orange for maybe 30 seconds

and then take the orange away.

110

: 00:05:22

And we'd look at these time series and

we're trying to use these time series data

111

: 00:05:25

to actually create these smell prints.

112

: 00:05:28

And that's a very difficult thing to do.

113

: 00:05:30

It actually most of the

time took machine learning.

114

: 00:05:32

So once again, this is maybe a

more advanced data analyst role.

115

: 00:05:35

'cause most data analyst roles.

116

: 00:05:37

You're not really using machine learning.

117

: 00:05:39

This type of machine learning is often

called classification, where you're

118

: 00:05:42

basically trying to match data to a

certain category based off of its data.

119

: 00:05:47

So for example, I could bring

an apple near it, right?

120

: 00:05:50

And the sensors would react.

121

: 00:05:52

Maybe they'd go all down, and if

I brought an orange next to it,

122

: 00:05:55

maybe all the sensors would go up.

123

: 00:05:56

And so you can come up with some sort of

an algorithm that would be like, okay,

124

: 00:05:58

if the sensors go up, it's an apple.

125

: 00:06:00

If they go down, it's an orange.

126

: 00:06:02

Now that's really oversimplifying

it because apples and oranges,

127

: 00:06:06

those are only two things that

exist in the universe, right?

128

: 00:06:08

There's like so many

different things that exist.

129

: 00:06:10

We were playing a little

bit bigger stakes.

130

: 00:06:13

You can think of it when

you go to uh, TSA line and.

131

: 00:06:16

And sometimes they, you know, swab you

and they're trying to see if you have

132

: 00:06:19

like any drugs or any bombs on you.

133

: 00:06:21

That was kind of the stakes that we were

playing with in some of our use cases.

134

: 00:06:24

So I would take this data that oftentimes,

you know, was time series based.

135

: 00:06:28

We usually had like 12 to 16 to

24 different sensors on there.

136

: 00:06:33

And I would try to make these

smell prints using classification

137

: 00:06:36

models in machine learning.

138

: 00:06:38

Now, a lot of the time I was

doing this in Python python's.

139

: 00:06:40

Great for doing things

in machine learning.

140

: 00:06:42

There was even some simple

algorithms that I created that were.

141

: 00:06:45

Based in Excel, but

they are pretty simple.

142

: 00:06:47

The more complicated stuff.

143

: 00:06:48

I was doing Python at the time.

144

: 00:06:50

Also, just because we were doing

a lot of these experiments, SQL

145

: 00:06:53

would've been really helpful.

146

: 00:06:54

We weren't actually using SQL

as much as we should have.

147

: 00:06:56

We really should have been using sql.

148

: 00:06:58

Uh, looking back on it a little bit more.

149

: 00:07:00

The third experience I wanna tell you

about was when I was doing my own,

150

: 00:07:03

uh, data science consultancy firm,

and I got hired by a cybersecurity

151

: 00:07:07

company to help them with a few things.

152

: 00:07:09

So obviously we live in this digital age.

153

: 00:07:11

Cybersecurity is really

important, so there's a lot of

154

: 00:07:15

opportunity in cybersecurity.

155

: 00:07:16

And the interesting thing

about cybersecurity is a

156

: 00:07:20

lot of the data is like.

157

: 00:07:21

Hidden in logs, because basically anything

you do online, anything you do on the

158

: 00:07:25

internet gets logged one way or another.

159

: 00:07:28

Like it's, it's in there.

160

: 00:07:29

They're capturing everything, but when

you capture everything, you're kind

161

: 00:07:33

of capturing nothing at the same time

because it's really hard to figure out

162

: 00:07:36

what's the signal amongst so much noise.

163

: 00:07:39

And so this company in particular

was basically getting a bunch of

164

: 00:07:42

internet logs for companies in what

you can consider their workspaces.

165

: 00:07:45

So for instance, all of their Microsoft

logs, all of their Google logs, if

166

: 00:07:49

they're using Slack, their Slack logs,

maybe their employee customer history.

167

: 00:07:54

Just think of like anything

a company might be interested

168

: 00:07:57

in from a cybersecurity stand.

169

: 00:07:58

We were just getting a bunch of the logs.

170

: 00:08:00

Now in these logs, there's maybe

little needles in the haystack.

171

: 00:08:04

There's maybe little gems

that can be pulled out.

172

: 00:08:06

It requires a lot of analysis to

try to figure out what's in there.

173

: 00:08:10

Just imagine you're getting

like a ton of hay and you have

174

: 00:08:13

to find this little needle.

175

: 00:08:14

And so my job was to go in there and try

to see if there was any needles, anything

176

: 00:08:17

that was like really worth diving into

and investigating more, and also just

177

: 00:08:20

summarizing everything that was happening.

178

: 00:08:22

This is how many logins

you had on Google today.

179

: 00:08:24

This is how many, you know,

logouts you had on Microsoft.

180

: 00:08:27

You know, this is how many users

you had from these different states.

181

: 00:08:30

Just like from these giant enterprise

organizations where they have thousands of

182

: 00:08:34

employees and a bunch of things going on.

183

: 00:08:36

Like how do you know

everything's going okay?

184

: 00:08:38

Are you sure that like everyone

is where they say they are?

185

: 00:08:41

Are you sure you don't have any intruders,

you know, people accessing stuff from

186

: 00:08:45

a place that you probably shouldn't?

187

: 00:08:46

Those types of things.

188

: 00:08:47

So we were basically taking.

189

: 00:08:48

These huge dumps of logs that weren't

really important, that weren't really

190

: 00:08:52

interesting, and aggregating them and

trying to find the interesting things.

191

: 00:08:56

And then also making sure

that nothing nefarious was

192

: 00:08:58

going on to do that analysis.

193

: 00:09:00

I was actually using all of Python, but I

could really choose what tool I wanted to.

194

: 00:09:03

I just chose Python personally because

I'm very comfortable in Python.

195

: 00:09:07

I'm, I'm decently good at Python, uh,

and I can do things quickly with Python.

196

: 00:09:11

I probably couldn't have done

this as easily, like in Excel.

197

: 00:09:15

You probably could have done similar

stuff in SQL if you wanted to.

198

: 00:09:18

One thing I really like about Python

is it can do anything, maybe not

199

: 00:09:21

extremely well, but it can do anything.

200

: 00:09:23

Um, so like I was doing all my analysis.

201

: 00:09:26

Uh, in Python and I was creating

data visualizations in Python.

202

: 00:09:28

They even used a lot of the insights

I found, like in terms of aggregates.

203

: 00:09:32

They basically like aggregated all of

their customers data and would publish

204

: 00:09:35

like a, a yearly or, or biannual

report of like cybersecurity incidents.

205

: 00:09:40

And so they were kind of like with graphs

that I was creating with some of these

206

: 00:09:43

KPIs or metrics that I was monitoring.

207

: 00:09:45

That way they could kind of inform

the cybersecurity, you know, fields

208

: 00:09:50

all of their customers about like

what the trends and what we were

209

: 00:09:52

seeing on a big picture standpoint.

210

: 00:09:54

And that was actually really useful 'cause

people would start to like read that and

211

: 00:09:57

be like, oh, I really like this company.

212

: 00:09:59

I wanna work with them.

213

: 00:10:00

And that would bring in new customers.

214

: 00:10:01

So even though like I was doing

that analysis for individual

215

: 00:10:05

customers at an individual level.

216

: 00:10:07

That analysis actually ended up being

really useful for their marketing

217

: 00:10:09

team as well to get more sales and

more customers in the pipeline.

218

: 00:10:12

Now, I've actually worked for way

more than just these three companies.

219

: 00:10:15

I've probably done work for

about 12, including like the Utah

220

: 00:10:18

Jazz, Harley Davidson, and some

other really big names like MIT.

221

: 00:10:23

If you want to hear more about

those, I'll be talking about

222

: 00:10:25

them more in my newsletter.

223

: 00:10:26

So you can

subscribe@datacareerjumpstart.com

224

: 00:10:30

slash newsletter, and

I'll be talking more.

225

: 00:10:32

About these experiences in the newsletter,

but if you want me to talk about it

226

: 00:10:35

on the podcast or YouTube as well,

let me know in the comments down below

227

: 00:10:38

and maybe I'll do some future episodes

on that if we get enough comments.

228

: 00:10:41

As always, thanks for watching

and I'll see you in the next one.