1
00:00:05,321 --> 00:00:15,561
Hi, my name is Yijia Liang. I'm the co-founder of Compound AI. We sell analysts for buy-side finance, and specifically we specialize in long-running analysts.

2
00:00:15,821 --> 00:00:21,921
Think cloud agents that Devon or Factory might do for coding agents, but tailoring that for buy-side institutions.

3
00:00:22,321 --> 00:00:25,021
I was previously an analyst in a previous life. I was

4
00:00:25,041 --> 00:00:38,321
a vice president at Code2 for about three years, specializing in late-stage private software investments. Spent way too much time in Excel, PowerPoint, and late nights doing that type of work and so decided to split and start building tools to replace my old job.

5
00:00:39,701 --> 00:00:43,561
Quick show of hands, how many people in here have used user instinct in the last day?

6
00:00:45,741 --> 00:00:49,501
I expected more hands, but maybe people are too shy to raise them.

7
00:00:50,421 --> 00:00:56,041
And how many people know an analyst in finance that fully delegates their work to AI, fully

8
00:00:56,041 --> 00:00:58,261
hands off, running it autonomously?

9
00:00:58,501 --> 00:01:00,301
One, but a lot fewer.

10
00:01:02,061 --> 00:01:06,241
I think if you think about the question, why did personal agents take off within the last

11
00:01:06,241 --> 00:01:09,521
three months, internally we came up with three main answers.

12
00:01:10,121 --> 00:01:12,701
The first, reliable background agents.

13
00:01:12,701 --> 00:01:17,761
You can delegate a task, book a restaurant, book a flight, and it'll do it reliably now.

14
00:01:18,101 --> 00:01:19,921
Computer use agents have gotten sophisticated.

15
00:01:20,521 --> 00:01:22,241
The models themselves have gotten better.

16
00:01:23,801 --> 00:01:25,101
The second is proactivity.

17
00:01:25,421 --> 00:01:26,961
It'll do things without you asking.

18
00:01:27,121 --> 00:01:30,281
It's part of that magical experience when instinct wakes you up saying,

19
00:01:30,441 --> 00:01:32,581
it's your wife's anniversary, go buy some flowers.

20
00:01:32,801 --> 00:01:34,561
Or better yet, I already bought the flowers for you.

21
00:01:35,461 --> 00:01:36,841
And finally, integrations.

22
00:01:37,081 --> 00:01:39,601
It's connected to the systems of context that matter most for you.

23
00:01:39,601 --> 00:01:41,961
Email, drive, calendar.

24
00:01:44,481 --> 00:01:51,261
all the factors that make personal agents work so well doesn't really translate to institutional

25
00:01:51,261 --> 00:01:51,861
finance today.

26
00:01:52,601 --> 00:01:57,521
If we think about background tasks, something as simple as booking a restaurant reservation,

27
00:01:58,441 --> 00:02:03,541
the analog equivalent in by-side finance might be how do you go from a data room to a fully

28
00:02:03,541 --> 00:02:06,561
prepared set of IC models, slides and memos.

29
00:02:06,561 --> 00:02:11,741
taking a process that would normally take four weeks, multiple rounds of iteration and review,

30
00:02:11,981 --> 00:02:13,661
and compressing that to a few days.

31
00:02:14,781 --> 00:02:20,741
If we think about proactivity, your managing director doesn't really want an agent rogue

32
00:02:20,741 --> 00:02:25,181
running around doing a bunch of things and burning tokens without them providing and soliciting input.

33
00:02:25,981 --> 00:02:30,561
And then finally, if you think about the context that matters, it's not just email, it's not just calendar.

34
00:02:30,941 --> 00:02:33,701
It could be a data room with 15,000 files.

35
00:02:33,701 --> 00:02:39,201
It could be SharePoint box with petabytes of unstructured data the firm has collected in its 40-year history.

36
00:02:39,421 --> 00:02:41,081
So how do you deal with these problems?

37
00:02:42,561 --> 00:02:50,481
So the purpose of today's talk is to discuss a bit more about how we at Compound have approached the problem of solving long-running agents in finance.

38
00:02:51,041 --> 00:02:56,121
On one hand, it's creating a real teammate that collaborates with you in the deal team.

39
00:02:56,281 --> 00:02:58,481
So that's multiplayer embedding that into the app.

40
00:02:59,101 --> 00:03:06,561
On the other hand, it's solving for how can you run long tasks in parallel in the background, reliably execute them.

41
00:03:07,181 --> 00:03:10,001
And so for that, we actually rebuilt Office from scratch.

42
00:03:10,661 --> 00:03:13,241
And then finally, it's how do you manage the context of everything.

43
00:03:13,441 --> 00:03:16,941
Everything we just discussed, petabytes of data in SharePoint in Box.

44
00:03:17,261 --> 00:03:20,141
And we'll talk a little bit about how we've designed a solution for that.

45
00:03:21,961 --> 00:03:26,541
So the first question we have to ask, I think, is why does multiplayer actually matter for finance?

46
00:03:26,541 --> 00:03:31,981
If you look at the anatomy on a deal team, the type of tasks that matter for long-running

47
00:03:31,981 --> 00:03:36,921
type of workflows, for example, going from a data room into all the sorts of materials

48
00:03:36,921 --> 00:03:41,201
that make its way in front of an investment committee, it doesn't work for the single-player

49
00:03:41,201 --> 00:03:42,021
siloed workflows.

50
00:03:42,741 --> 00:03:48,821
Today, for people who have deployed Cloud or ChatGPT in their institutions, it's all single-player.

51
00:03:49,161 --> 00:03:55,301
People working in a plug-in, people working in the web app, producing outputs, but invariably,

52
00:03:55,301 --> 00:03:57,581
those people need to collaborate with their team.

53
00:03:57,861 --> 00:03:58,601
And what does that look like?

54
00:03:58,741 --> 00:04:01,441
It means downloading the file, uploading it into email,

55
00:04:01,621 --> 00:04:03,301
uploading it to Teams, asking for feedback.

56
00:04:04,641 --> 00:04:06,141
I think the right way to approach this,

57
00:04:06,181 --> 00:04:07,941
and this is the way we've thought about at Compound,

58
00:04:08,121 --> 00:04:10,681
is how do you get every person who matters on a deal

59
00:04:10,681 --> 00:04:11,641
onto the same thread?

60
00:04:11,821 --> 00:04:13,621
How can they all talk to the same agent

61
00:04:13,621 --> 00:04:16,961
so that one agent can share the context of everyone who matters

62
00:04:16,961 --> 00:04:18,541
for any particular work stream?

63
00:04:20,701 --> 00:04:23,581
When we thought about how do we actually bring multiplayer

64
00:04:23,581 --> 00:04:24,821
into a familiar context,

65
00:04:25,301 --> 00:04:27,081
We looked again towards personal agents.

66
00:04:27,501 --> 00:04:28,521
Why did they work so well?

67
00:04:28,901 --> 00:04:30,781
Well, you can easily text one in iMessage,

68
00:04:30,901 --> 00:04:32,021
or you can text one in WhatsApp.

69
00:04:33,141 --> 00:04:35,501
So for agents at Compound, every single one you create

70
00:04:35,501 --> 00:04:37,041
comes with a built-in inbox.

71
00:04:37,761 --> 00:04:39,141
Because ultimately, at the end of the day,

72
00:04:39,321 --> 00:04:41,621
for institutional finance, most workflows run through Outlook.

73
00:04:41,941 --> 00:04:44,401
You'll have chains of emails, 20 or more running.

74
00:04:44,961 --> 00:04:47,041
You want an agent to be part of every single one

75
00:04:47,041 --> 00:04:47,841
of those conversations.

76
00:04:50,481 --> 00:04:53,661
Designing email for agents is actually non-trivial.

77
00:04:53,661 --> 00:04:56,441
It's not just as simple as hooking it up with an email inbox.

78
00:04:56,721 --> 00:05:02,681
It requires A, authenticating, making sure everyone part of that thread has read and write access to that agent.

79
00:05:03,921 --> 00:05:07,241
Ultimately, at these financial firms, privacy is incredibly important.

80
00:05:07,441 --> 00:05:09,701
You're always under NDA no matter what deal you look at,

81
00:05:09,841 --> 00:05:12,921
and so you want to make sure everyone on that thread can actually talk to the agent.

82
00:05:13,401 --> 00:05:17,901
But then beyond authentication, you also have to think about how do you design a real human-like experience,

83
00:05:18,181 --> 00:05:20,101
because after all, we're trying to create a real teammate.

84
00:05:20,101 --> 00:05:25,521
If multiple people email in the same thread in quick succession, you don't want multiple

85
00:05:25,521 --> 00:05:27,121
acknowledgments spamming that group.

86
00:05:27,301 --> 00:05:31,001
You actually want to bake in some more latency to make it feel more human.

87
00:05:31,201 --> 00:05:34,741
And so there are other design decisions when thinking about how to create real email agents

88
00:05:34,741 --> 00:05:36,101
for enterprise finance.

89
00:05:37,821 --> 00:05:42,581
But multiplayer aside, we realized even after you solicit the input of everyone who matters,

90
00:05:43,541 --> 00:05:48,161
you wanted to scale up the number of tasks and run them all in parallel, existing off

91
00:05:48,161 --> 00:05:49,781
the shelf solution simply don't cut it.

92
00:05:50,181 --> 00:05:51,461
We evaluated a lot.

93
00:05:51,841 --> 00:05:54,421
The desktop plug-ins of Excel, PowerPoint, and Word.

94
00:05:54,841 --> 00:06:00,001
We evaluated computer use using virtual machines to scale up the number of agents that could

95
00:06:00,001 --> 00:06:02,241
work in Excel, PowerPoint, and Word natively.

96
00:06:02,601 --> 00:06:06,401
And we also evaluated a ton of off-the-shelf engines, whether those are open source packages

97
00:06:06,401 --> 00:06:10,501
like openpyxl or commercial engines that are available for purchase.

98
00:06:11,061 --> 00:06:13,701
None of them really cut it for what we were trying to look at.

99
00:06:13,701 --> 00:06:19,381
And with coding agents dramatically improving and bringing the cost of software down, we

100
00:06:19,381 --> 00:06:21,361
decided to just build our own office from scratch.

101
00:06:22,661 --> 00:06:28,441
And like a great AI engineer would, we ran agents in many loops, spent 1.5 trillion input

102
00:06:28,441 --> 00:06:32,241
output tokens and over 8 million lines of code to rewrite office from scratch.

103
00:06:32,721 --> 00:06:35,101
And I'll walk through a little bit of how we approached it.

104
00:06:36,001 --> 00:06:40,901
it's as simple as pointing agents at an objective function, which is recreate office

105
00:06:40,901 --> 00:06:42,901
with full fidelity and then letting it run.

106
00:06:44,541 --> 00:06:46,021
take Excel as an example.

107
00:06:46,781 --> 00:06:49,421
We started with an open source solution called IronCalc.

108
00:06:49,541 --> 00:06:54,761
We decided, okay, how can you recreate this perfectly for Excel, use Office as an Oracle,

109
00:06:55,141 --> 00:07:00,201
and just run loops, benchmarking it on real files from both real world corpus that we

110
00:07:00,201 --> 00:07:03,001
grabbed online and also ones that internally we created.

111
00:07:03,461 --> 00:07:10,301
And 800,000 lines of rust later, we had a much more performant compared to Excel, one

112
00:07:10,301 --> 00:07:11,901
one for one replica of the product.

113
00:07:12,961 --> 00:07:16,601
With this, we're actually able to do that for not only Excel,

114
00:07:16,761 --> 00:07:17,941
but also PowerPoint and Word.

115
00:07:18,521 --> 00:07:21,161
So we ran loops across every single one of these products,

116
00:07:21,281 --> 00:07:22,561
and we have faithful recreations.

117
00:07:22,741 --> 00:07:25,521
You can try it for yourself today at get compound.ai.

118
00:07:25,801 --> 00:07:28,081
And it's the same method, loop agents

119
00:07:28,081 --> 00:07:31,401
against Office as a judge, using rust renders

120
00:07:31,401 --> 00:07:34,541
to make sure that the rendering of these products

121
00:07:34,541 --> 00:07:35,401
was pixel perfect.

122
00:07:35,721 --> 00:07:38,541
So what you see in Compound is what

123
00:07:38,541 --> 00:07:39,821
you see when you download to Office.

124
00:07:39,821 --> 00:07:45,961
We needed one-for-one parity because at the end of the day, these firms only trust what they can download and what they can send.

125
00:07:48,001 --> 00:07:49,761
Now, let's talk about performance.

126
00:07:50,041 --> 00:07:55,221
Why is a headless engine so much better compared to using a VM or using desktop Excel?

127
00:07:55,481 --> 00:08:00,141
Well, for one, desktop Excel or PowerPoint or Word is bound to that CPU.

128
00:08:00,661 --> 00:08:04,001
With headless engines, you can do tasks in parallel much, much faster.

129
00:08:04,381 --> 00:08:07,741
You can build out a complex model in a much shorter amount of time.

130
00:08:07,741 --> 00:08:11,641
You can recalculate 20 megabyte Excel's 4x faster.

131
00:08:12,021 --> 00:08:16,481
So every single piece of work an analyst would expect an agent to do,

132
00:08:16,641 --> 00:08:21,301
because, again, their expectation is quality of a human at much lower speed and much lower cost.

133
00:08:21,601 --> 00:08:23,901
You have to have a headless engine to do this.

134
00:08:24,221 --> 00:08:29,121
If you tried to scale this up with Excel and PowerPoint, I think you'd be going down the wrong path.

135
00:08:30,801 --> 00:08:32,321
But it's not just about performance.

136
00:08:32,961 --> 00:08:36,301
Owning the editor actually allows you to create a much richer product experience

137
00:08:36,301 --> 00:08:39,781
that allows you to work and collaborate with agents asynchronously

138
00:08:39,781 --> 00:08:42,261
much better than you can with other products.

139
00:08:42,641 --> 00:08:45,161
So this entire slide deck you're seeing today,

140
00:08:45,421 --> 00:08:48,681
I built in Compound just by dropping in comments.

141
00:08:48,921 --> 00:08:50,461
I had it create a first draft,

142
00:08:50,481 --> 00:08:52,241
and as you can see here on the UI,

143
00:08:52,461 --> 00:08:54,781
I dropped in comments for every single piece of formatting.

144
00:08:55,001 --> 00:08:57,301
For my consulting days, for my co2 days,

145
00:08:57,461 --> 00:08:59,481
all the learnings I had on what makes a great slide,

146
00:08:59,641 --> 00:09:01,241
I imparted with the agent.

147
00:09:01,921 --> 00:09:03,301
This can also work for multiplayer.

148
00:09:03,581 --> 00:09:05,421
You can have multiple people part of the same team

149
00:09:05,421 --> 00:09:09,101
using the same multiplayer infrastructure we set up, dropping in comments.

150
00:09:09,441 --> 00:09:13,281
So for a team of five, you can rip through a 100-page deck in an afternoon

151
00:09:13,281 --> 00:09:16,661
rather than sitting in a room revising it for multiple weeks.

152
00:09:18,681 --> 00:09:20,901
Finally, with respect to building your own engine,

153
00:09:21,081 --> 00:09:24,101
it also allows us to run iterative checking loops much more as well.

154
00:09:24,401 --> 00:09:28,661
So if you have a task that, for example, is writing a complex model,

155
00:09:28,941 --> 00:09:30,161
it might take a few hours.

156
00:09:30,841 --> 00:09:33,801
But how can we make that faster? How can we make it more accurate?

157
00:09:33,801 --> 00:09:39,021
If you are using an openpyxl code library, there is no render you can use to check formatting.

158
00:09:39,421 --> 00:09:42,901
You are just guessing and hoping that the model writes the right thing and maybe you

159
00:09:42,901 --> 00:09:46,481
do some parsing to see a cell as a particular color or has a particular font.

160
00:09:46,861 --> 00:09:51,461
If you try to use Microsoft's own interface, you are going to be limited by what Microsoft

161
00:09:51,461 --> 00:09:53,701
provides you, which is admittedly very little.

162
00:09:54,541 --> 00:09:59,181
With an own engine, you actually own the write, render, check and fix loop.

163
00:09:59,181 --> 00:10:02,801
so you can rapidly iterate on the output for the user.

164
00:10:04,881 --> 00:10:06,921
So the last thing I sort of want to discuss today

165
00:10:06,921 --> 00:10:09,581
beyond multiplayer and building our own office

166
00:10:09,581 --> 00:10:12,421
is how to deal with the context that these firms have.

167
00:10:15,181 --> 00:10:16,641
For firms on the buy side,

168
00:10:16,861 --> 00:10:20,581
they deal with situations where it crosses both breadth and depth of data.

169
00:10:20,861 --> 00:10:24,501
You may have situations where you have a data room of over 15,000 files,

170
00:10:24,681 --> 00:10:25,021
and

171
00:10:25,101 --> 00:10:27,861
then files in that data room might be well over 20 megabytes.

172
00:10:28,581 --> 00:10:38,861
People in finance complain all the time, hey, I have an Excel model that's 30 megabytes, it takes five minutes to recalculate because we have two million formulas and five sensitivity tables that need to recalculate every single time.

173
00:10:39,341 --> 00:10:40,461
So how do you deal with this?

174
00:10:41,641 --> 00:10:48,321
Well, for breadth, I think, again, we take a look at a page out of the coding book and apply agent swarms to the large number of files.

175
00:10:48,321 --> 00:10:53,481
files, no longer do you have to index everything in a vector DB unless it's for a very specific

176
00:10:53,481 --> 00:10:53,941
use case.

177
00:10:54,181 --> 00:10:58,961
It's actually much preferable to use coding agents to read files, either through grep

178
00:10:58,961 --> 00:11:04,101
or through some other methodology, and then have a lead agent slice the contents of the

179
00:11:04,101 --> 00:11:07,841
file directory and then have swarms of agents come back to you with insights.

180
00:11:08,141 --> 00:11:12,581
And then you iterate that in a loop until you're satisfied with the output required.

181
00:11:13,701 --> 00:11:17,181
Furthermore, through this process, you're actually able to ground every single finding

182
00:11:17,181 --> 00:11:21,621
in citations in the source document so that you know exactly what you show on the screen

183
00:11:21,621 --> 00:11:24,541
is linked to some number that the analyst can trust.

184
00:11:26,441 --> 00:11:32,421
If we think about the depth of data, take that 30 megabyte workbook, how can you actually

185
00:11:32,421 --> 00:11:33,821
process this more accurately?

186
00:11:34,601 --> 00:11:39,101
In today's world, if you were to just use ChatGPT or Claude off the shelf to process

187
00:11:39,101 --> 00:11:43,321
that 30 megabyte model, it would write code, it would use open source packages like open

188
00:11:43,321 --> 00:11:48,401
Excel, it would be expensive and slow because it's printing out every single piece of context

189
00:11:49,091 --> 00:11:51,541
about that model in code for the model to read.

190
00:11:52,161 --> 00:11:56,121
But because we've developed our own engine for Excel, we're actually not only more accurate

191
00:11:56,121 --> 00:12:01,481
but a lot faster and cheaper because the engine processes it and then lays it out in a very

192
00:12:01,481 --> 00:12:05,041
easy to read and easy to process format for the model.

193
00:12:05,721 --> 00:12:11,021
This means we can expose a much richer set of APIs and a much richer set of tools for

194
00:12:11,021 --> 00:12:11,901
the agent to work over.

195
00:12:11,901 --> 00:12:16,141
And that allows us to process these models in a much faster manner.

196
00:12:17,721 --> 00:12:22,141
So if we sort of take a step back and think about, okay, what is all of this context built

197
00:12:22,141 --> 00:12:22,401
for?

198
00:12:23,261 --> 00:12:27,581
At the end of the day, what we want to do is prevent people from spending time finding

199
00:12:27,581 --> 00:12:30,821
files, from reminding the model this is how to do things.

200
00:12:31,161 --> 00:12:34,621
Instead, they should be focused on imparting their judgment on the deal process.

201
00:12:34,821 --> 00:12:40,181
They should be focused on tailoring, training agents, and deploying them on every part of

202
00:12:40,181 --> 00:12:40,921
the deal flow.

203
00:12:40,921 --> 00:12:47,821
And so it goes from the bottom level, connected sources, off-the-shelf connectors to SharePoint or Outlook,

204
00:12:48,061 --> 00:12:51,141
to the cited findings grounded in the sources that matter.

205
00:12:51,441 --> 00:12:55,121
And then finally, embedding in the preferences, company knowledge that matters.

206
00:12:55,401 --> 00:12:58,421
How do you do a model? How should it look? Which differs by every single firm.

207
00:12:58,641 --> 00:13:02,941
So that, for humans, they spend most of their time imparting judgment.

208
00:13:03,101 --> 00:13:06,121
They spend most of their time thinking about, is this actually a good deal?

209
00:13:06,441 --> 00:13:09,241
Not, did all the numbers tie? Does the model look right?

210
00:13:09,241 --> 00:13:11,801
will my VP yell at me because it looks a different way.

211
00:13:14,881 --> 00:13:17,801
So, oh, Kirscher needs to update.

212
00:13:20,561 --> 00:13:25,721
If we take a step back and think about the type of tasks that currently are on autopilot

213
00:13:27,281 --> 00:13:31,501
versus what are on copilot, I would say there's already several finance tasks.

214
00:13:31,621 --> 00:13:33,261
You can do pretty well with off-the-shelf tools.

215
00:13:33,621 --> 00:13:38,161
Spreading comps, as the chat GPT head of finance mentioned earlier, they have now their own

216
00:13:38,161 --> 00:13:42,281
products for this, but I think there are a lot of longer running tasks that are not capable

217
00:13:42,281 --> 00:13:43,361
with existing solutions.

218
00:13:43,761 --> 00:13:47,301
How do you draft a 50-page IC presentation off the cuff?

219
00:13:47,521 --> 00:13:53,741
How do you get agents to build out a model exactly the way you want it for a private

220
00:13:53,741 --> 00:13:55,921
equity professional or a hedge fund professional?

221
00:13:56,401 --> 00:14:01,901
I think those are tasks that with off-the-shelf tools still exist within the co-pilot realm.

222
00:14:02,001 --> 00:14:06,301
People want to be embedded deeply in that process without letting go.

223
00:14:07,281 --> 00:14:10,941
But with some of the tools we've built today, I think we're able to expand the aperture

224
00:14:10,941 --> 00:14:14,401
for which Autopilot can take over in finance firms.

225
00:14:15,141 --> 00:14:20,921
So rather than just using tools embedded in Excel as plugins, hand-holding along the way,

226
00:14:21,121 --> 00:14:22,361
you can start delegating.

227
00:14:22,561 --> 00:14:26,001
You can start becoming more ambitious about what AI is capable of.

228
00:14:26,221 --> 00:14:30,481
If I think about where we are today in the world of finance, it very much reminds me

229
00:14:30,481 --> 00:14:38,421
of when Opus 4.5 launched last December, where people stopped using code as a simple co-pilot

230
00:14:38,421 --> 00:14:43,121
to generate the next file or the next line of code, but they started delegating far more

231
00:14:43,121 --> 00:14:48,621
ambitious tasks. They started realizing, maybe I can just truly vibe it. If I have sufficient

232
00:14:48,621 --> 00:14:54,821
checks, if I have sufficient tests, maybe I can just trust all the code it writes. I think

233
00:14:54,821 --> 00:14:59,021
with some of the tooling we've built today, along with what we're planning in the future,

234
00:14:59,021 --> 00:15:02,921
these are the type of problems you have to solve before you get to that point for finance.

235
00:15:03,521 --> 00:15:07,621
How can analysts, how can the entire deal team become more ambitious with AI?

236
00:15:07,981 --> 00:15:12,081
How can they just drop in a data room and decide, okay, I'm going to trust every single one of the outputs.

237
00:15:12,341 --> 00:15:16,961
And then let's do that tenfold so we can see more deals, so we can expand the aperture of what matters.

238
00:15:18,801 --> 00:15:21,961
So to conclude, enterprise agents are certainly coming to finance.

239
00:15:22,761 --> 00:15:25,781
I strongly believe we're at an inflection point in the adoption curve

240
00:15:25,781 --> 00:15:28,001
and that people are slowly becoming more ambitious

241
00:15:28,001 --> 00:15:30,261
and they're moving away from co-pilots to autopilots.

242
00:15:30,781 --> 00:15:34,021
And so this was sort of how we've approached it here at Compound.

243
00:15:34,681 --> 00:15:37,841
If you work at a firm and you're interested in deploying our agents

244
00:15:37,841 --> 00:15:40,341
to drive more of your deal work, please reach out.

245
00:15:40,501 --> 00:15:43,401
Or if you're an engineer and you're curious about talking about our approach

246
00:15:43,401 --> 00:15:45,961
and sharing ideas, please either reach out online

247
00:15:45,961 --> 00:15:47,381
or I'll be around if you want to find me.

248
00:15:47,621 --> 00:15:48,221
Thanks so much.