US – Hyperautomating the Customer Service Experience with Gen AI and Autonom8

Renga: Alright! That’s a wrap for the song. Where to start to our webinar. Welcome everyone. Thanks for joining us and making time for this webinar. I promise this would be 60 min well spent today we had second occurrence of the webinar with Autonom8, where we showcase, our platform and what it could do for automating customer service with

Renga: A8tudio engine AI joining us today is our host for the evening. Ranjit Pradhmanavin, who’s also our co-founder and CTO. He will be taking us through this webinar. Please feel free to put in your questions at the QA. Section, and I will be orchestrating the Christians and transit responses, as we take it. Towards the end we will leave ample time for questions, so keep them flowing, and we will compile it so that Ranjit can answer each one of you as we go along.

Ranjit: Alright over to you, Ranjit, all the best. And let’s get started. Yeah, thanks, Ranga, and and welcome everyone. Good morning. Good afternoon. Good evening. Wherever you are.

Ranjit: So first of all, let me check. Is my screen visible.

Renga: Yes, it’s visible. Thank you.

Ranjit: Great. So so thanks again to all of you for making the time to come and join the webinar. This as as Ranga said, we’ll spend about 50 min on this webinar. Leave some time for questions the the approach is going to be, you know, I’ll start with a few slides describing what we do at autonomate. Which kind of sets the context. And then we are going to specifically focus on using Jenny. I in in, you know, for customer service, and to enhance the customer service experience.

Ranjit: So you know, a short, a few short slides in the beginning, followed by a bunch of different demos. So let me start by describing what we do at autonomate.

Ranjit: Our logo is enabling the autonomous enterprise. And the idea here is that in in very much the same way that you saw the evolution of autonomous vehicles where they had intelligence, the ability to make decisions, the ability for a human in the loop, and all of these kinds of things we believe many of those same concepts are applicable to enterprises as well. So we think of that as the autonomous enterprise.

75
00:09:11.760 –> 00:09:40.320
Ranjit: So just to describe our our platform in in a bit more detail we so II first let me define the word platform. So what we mean by the platform that we sell at autonomate is that we. We give you the enterprise, a low code studio, along with a collection of components with which you can assemble your own workflows. And actually, we have a more broad perspective on workflows. We think of them as customer journeys.

76
00:09:40.320 –> 00:09:49.549
Ranjit: So, for example, if if you’re a banking institution, you can think of a customer journey being some where someone comes and applies for a loan, and

77
00:09:49.570 –> 00:10:05.750
Ranjit: where where the loan gets processed in the back end and a decision gets sent back to the customer. It’s a full end to end customer experience in the context of customer support. This may be a situation where someone comes in to ask for support on a particular product, for example.

78
00:10:06.040 –> 00:10:17.980
Ranjit: And this information then goes to the back end. You may have an agent assisting you, or you may have the whole process completely automated. And eventually the customer goes back with a response to their question.

79
00:10:18.410 –> 00:10:33.169
Ranjit: So if you look at you know the the model of a customer journey? It generally starts with a customer approaching the enterprise. They can. They can approach you through any channel which could be virtual or it could be live. And we we call that the intake process.

80
00:10:33.230 –> 00:10:54.589
Ranjit: We have low code modules that allow you to create these intake experiences either via chat or via apps. And that’s where most of my focus is gonna be today. But just to give you a full picture, once the information comes in through these channels. You have a back end business process that kicks into a gear once the information comes in.

81
00:10:54.590 –> 00:11:10.250
Ranjit: and this consists of multiple people working together, integrating with different systems like Crm systems, ticketing systems, and so on and resolving the the customer inquiry or question or request.

82
00:11:10.670 –> 00:11:18.630
Ranjit: Now, while these customer journeys are going on, we also have a real time analytics. Capability that is looking continuously at

83
00:11:18.890 –> 00:11:36.430
Ranjit: the the business process or the customer journey, and giving you insights and recommendations on on where this can be improved in the future. It comes with a few important capabilities like detecting sentiment. You know, identifying anomalies and and things of that nature.

84
00:11:37.950 –> 00:11:41.710
Ranjit: So let me now kind of drill down into how a

85
00:11:42.080 –> 00:11:51.969
Ranjit: our platform is used broadly to solve. Bpo use cases. Now, Bpo is a broad term. And and really, what we want to do is focus on automated customer support.

86
00:11:52.410 –> 00:12:07.539
Ranjit: So if you look at the benefits to the customer you, we, we provide conversational AI for personalized interactions. And in this, you know, as the topic suggests, we’re going to talk about how we can use Jenny I. To create these conversational experiences.

87
00:12:07.630 –> 00:12:14.709
Ranjit: We we enable customers to come in through all different channels with support for text and voice. Separately.

88
00:12:14.740 –> 00:12:33.600
Ranjit: The same kind of capabilities is is applicable to helping agents. You can help them during the call by providing whispers or or recommendations on how they should respond to a particular customer. And finally, there’s capabilities. Once upon completion of the call to summarize it and store it, and learn from it.

89
00:12:34.360 –> 00:12:51.040
Ranjit: And then, if you take a step back and look at the benefits to the enterprise as a whole, you have, you know, a a whole set of integrations with with workflows and the ability to to extract knowledge from customer experiences and share them within your organization.

90
00:12:52.350 –> 00:13:02.280
Ranjit: So I said, we had a platform and I spoke. I showed you kind of the model for a customer journey. Our platform actually consists of these 6 modules if you will.

91
00:13:02.330 –> 00:13:16.060
Ranjit: So one way to think about it is like Microsoft Office, you have a collection of different modules that independently do very useful tasks, but are even more powerful when integrated together. So that same principle applies here.

92
00:13:16.360 –> 00:13:44.459
Ranjit: We have a a chat platform that enables you with a low code interface to create chats and dialogues. And during this process in a incorporate Llm’s and Gpt. And all of these new technologies as well, we have a a capability for a Ui to build and create uis for app experiences, contact center. And on the back end we have a collection of low code tools to create these end, to end workflows.

93
00:13:44.530 –> 00:13:47.059
Ranjit: A. 8 IQ is our analytics engine.

94
00:13:47.180 –> 00:13:59.189
Ranjit: And and it’s sitting at the bottom of everything we do so anytime a user interacts with an enterprise that information goes into our engine and we’re able to actually make use of it and learn from it.

95
00:13:59.740 –> 00:14:01.929
Ranjit: As I said today, I’m going to focus

96
00:14:02.190 –> 00:14:08.390
Ranjit: on the front end conversational platform and the applicability of geniii in this situation.

97
00:14:09.670 –> 00:14:20.630
Ranjit: Now, everyone’s been using genii. You know there’s lots of Apis available. And so you have many teams of developers who’ve been using the Apis and getting great results, and all of that.

98
00:14:20.840 –> 00:14:36.520
Ranjit: But you need something more to make this truly an enterprise class product, and these are the features that come that that we bundle together as part of our product for any enterprise that’s seeking to automate their customer support with genii.

99
00:14:37.150 –> 00:14:38.350
Ranjit: for example.

100
00:14:38.480 –> 00:15:05.240
Ranjit: quite often, most of the time, geniii or Gp. Let’s take Gpt. For example, will give you good results, accurate results. But not all the time. There’s issues like hallucinations. There’s issues like lack of, you know, staleness of information and things of that sort. So what you want is the ability. Anytime you get a response from a back end model like Gpt to have the ability to curate it and and fix it so that the better

101
00:15:05.320 –> 00:15:08.040
Ranjit: response can be used in the future.

102
00:15:09.060 –> 00:15:22.590
Ranjit: As with all conversational experiences, we provide kind of a personalized contextual capability. So you don’t have to repeat yourself, and I’ll show you all of these things in the demos. They’re seamless into escalation and de-escalation to support agents.

103
00:15:22.630 –> 00:15:31.839
Ranjit: the the whole chat. Conversational experience can be customized to your content. This could be structured, content, unstructured content. And again, I’ll show you Demos of this.

104
00:15:32.380 –> 00:15:35.179
Ranjit: Now on the infrastructure side.

105
00:15:35.240 –> 00:15:57.379
Ranjit: Anytime you use a back end model. You’re incurring an incremental cost, and these costs can add up really quickly. So what we try and do is provide you with the capability to manage your costs. And we do this in a in a few different ways, like giving you access to different models that you can use, based on the specific use case that you have and a and do a trade off between

106
00:15:57.380 –> 00:16:10.180
Ranjit: cost. And and you know the quality. So, for example, you might get almost as good quality with a different model. That might be 10% of the cost we give. We give you all of that wisdom in our Llm management console.

107
00:16:10.230 –> 00:16:32.890
Ranjit: There’s a a rule rule based capabilities for you to create these dialogues and conversations. So they don’t always have to use a A, A model. You can always, you can sometimes divert and use traditional Nlp and traditional techniques to respond to user questions. And then these are supported with all sorts of guardrails for security, accuracy, scope, and so on.

108
00:16:33.280 –> 00:16:47.980
Ranjit: I’m gonna jump into Demos now and by the way, we have, the way we have this set up is, you know, we’ll have the questions at the end, so please hold off on your questions, or start typing them into the QA. Box. And I’m happy. I’m happy to answer them at the end.

109
00:16:48.580 –> 00:16:54.059
Ranjit: So I have a few different demos showcasing Gen. AI capabilities as part of our platform.

110
00:16:54.260 –> 00:16:58.119
Ranjit: and the first one I’m going to start with is called the Travel Assistant.

111
00:16:58.140 –> 00:17:10.780
Ranjit: The Travel Assistant is looks at a collection of unstructured data for this demo. We have actually used the data from the website. Lonely planet. Lonely planet has about 300,000 pages on the web.

112
00:17:10.890 –> 00:17:14.680
Ranjit: And as you know, this is unstructured data. But it is public data.

113
00:17:14.869 –> 00:17:22.790
Ranjit: The objective of this travel assistant is is to give informational responses to people who are searching for travel information

114
00:17:23.410 –> 00:17:25.840
Ranjit: for this particular use case. We have

115
00:17:25.960 –> 00:17:42.979
Ranjit: found that oh, GPT. 3.5 is a very suitable model, the right price point. And for those of you who care about some of the technicalities under the hood. We’re using an approach called retrieval augmented generation. And I’ll explain that as we move forward through the demo

116
00:17:43.500 –> 00:17:58.049
Ranjit: the key highlight here is that due to our partnership with Google. We, we run in in Google Cloud by default. We have a partnership with them. And so we are able to take. We don’t scrape any site anytime. You want us to actually

117
00:17:58.120 –> 00:18:10.389
Ranjit: create a chat bot for any given website. We can configure that in less than an hour assuming that the website is a public website. So now I’m gonna jump and just show you, you know this thing in action.

118
00:18:11.360 –> 00:18:14.240
Ranjit: what you see, on the right side of your screen is R.

119
00:18:14.650 –> 00:18:18.379
Ranjit: Demo console, and I’m going to start off the travel assistant

120
00:18:18.780 –> 00:18:25.350
Renga: if you wouldn’t mind keeping them side by side, so that the right window is almost

121
00:18:25.560 –> 00:18:27.860
Renga: touching the right.

122
00:18:28.100 –> 00:18:39.249
Ranjit: Yeah, I can do that. I have a few other windows that get hidden behind Ranga. So anyway. So I’ll take care of it. Yeah.

123
00:18:39.470 –> 00:18:41.010
Ranjit: So

124
00:18:41.080 –> 00:18:48.430
Ranjit: okay. So let’s start with the first question. So the first thing is that the travel assistant introduces itself.

125
00:18:48.520 –> 00:18:55.289
Ranjit: and I’m going to type in a question that says, What’s the best time of year to visit Italy. Now, here’s what’s going on

126
00:18:55.840 –> 00:18:57.950
Ranjit: for any time. I type in a query.

127
00:18:58.130 –> 00:19:03.189
Ranjit: we are actually looking at the index of the lonely planet site.

128
00:19:03.260 –> 00:19:05.019
Ranjit: Just the same way Google does

129
00:19:05.350 –> 00:19:10.080
Ranjit: pulling out information that is specific to the question that’s being asked.

130
00:19:10.100 –> 00:19:11.839
Ranjit: assembling that together

131
00:19:11.860 –> 00:19:29.970
Ranjit: and then sending that to to the Gpt model to collate as sending that kind of as a prompt. If, if you’re familiar with that kind of approach, we create a prompt send that to the backend model. The the model assembles the information and presents that in a very user friendly form

132
00:19:30.020 –> 00:19:31.500
Ranjit: back to the

133
00:19:31.750 –> 00:19:34.989
Ranjit: back to the user. So there’s a

134
00:19:35.110 –> 00:19:41.720
Ranjit: there’s 2 things going on here. The first thing is that we are constraining the contextual information

135
00:19:42.270 –> 00:19:54.289
Ranjit: that is sent to the backend model with authoritative information that’s taken from the website. So this really minimizes the possibility of hallucinations and incorrect information.

136
00:19:54.460 –> 00:19:59.329
Ranjit: The second part is that for every result we can actually give you a link back

137
00:19:59.720 –> 00:20:08.209
Ranjit: to the page on the site which provided the maximum value information that went into this response.

138
00:20:09.210 –> 00:20:11.220
Ranjit: So let me ask a follow-up question.

139
00:20:15.670 –> 00:20:31.299
Ranjit: So the follow-up question is, Do I need a visa. And in this case, obviously, you know, when in in a conversational mode, you know that we’re talking. The context is Italy. So you know broadly the question was on to do. I need a visa to visit Italy, and and you’ll notice that

140
00:20:31.320 –> 00:20:32.680
Ranjit: you know the

141
00:20:32.710 –> 00:20:37.800
Ranjit: conversationally, I was able to understand that and give you a response

142
00:20:37.860 –> 00:20:40.000
Ranjit: on on that specific point.

143
00:20:41.160 –> 00:20:58.580
Ranjit: But now so Italy. So I use the topic of Italy. But to to provide context between the first 2 questions. The next question I’m I’m going to ask is Anne to visit Morocco? I’m actually switching topics here from the topic of Italy to the topic of visas, and, as you can see.

144
00:20:58.750 –> 00:21:11.430
Ranjit: the the system is able to kind of follow the trend in the conversation, and, you know, provide information about, you know, visiting Morocco at the visa requirements for visiting Morocco.

145
00:21:11.720 –> 00:21:18.690
Ranjit: I’m going to switch gears now and ask, so you can think of these as being somewhat general questions. I’m going to switch gears and ask,

146
00:21:19.350 –> 00:21:22.180
a multi part very specific question.

147
00:21:22.270 –> 00:21:30.060
Ranjit: The question is about the Lalbagh Botanical Gardens. This is an attraction in India. And I’ve asked a bunch of different questions

148
00:21:30.230 –> 00:21:40.589
Ranjit: normally in in any Q&A type environment. When you ask multiple questions. You have different pieces of different places from where this information needs to be pulled together. And that’s what

149
00:21:40.750 –> 00:21:46.159
Ranjit: that is. One of the capabilities that this rag technique gives you. It’s the ability to

150
00:21:46.440 –> 00:21:51.819
Ranjit: pull information from different places and put them together into intour response.

151
00:21:52.900 –> 00:21:55.290
Ranjit: I’m going to switch gears here now.

152
00:21:55.610 –> 00:22:00.559
Ranjit: And I’m going. I’m going to type in a question which is which was specifically.

153
00:22:01.530 –> 00:22:05.870
Ranjit: which was specifically found from the website of

154
00:22:07.640 –> 00:22:16.840
Ranjit: Of a lonely planet. This was a a customer complaining. Now I brought up another window on the left side of my screen, which is our

155
00:22:16.910 –> 00:22:21.230
Ranjit: live agent window. So this is, think of it as a text based contact center.

156
00:22:21.690 –> 00:22:27.860
Ranjit: So I’m typing in the question. Your website gave me incorrect advice. And now I am stuck in Tunisia.

157
00:22:28.120 –> 00:22:50.029
Ranjit: now, what what’s happening here is that the the the chat is able to detect that there is negative sentiment associated with this particular question, and a and then based on that, there’s a rule that says I, wanna a transfer me to to an agent. So now, rather than this, this person conversing with.

158
00:22:50.260 –> 00:22:57.799
Ranjit: The the virtual agent. They’re actually conversing with the live agent. So this agent can say, Hi! And you can see

159
00:22:57.900 –> 00:23:04.110
Ranjit: the information you know. Show up here. So you know, as you can see, there’s a conversation going back and forth.

160
00:23:04.350 –> 00:23:16.990
Ranjit: you know. I won’t dwell on this, but I just wanted to show you the ability to escalate to a live agent, and then the agent can complete the chat, and then that switches back to. you know, to the virtual agent.

161
00:23:17.520 –> 00:23:35.210
Ranjit: So I’m going to show you a couple of different things that kind of should clarify the power that that we can provide through this, everything is not just GPT. Based. You can, you can do certain things that are rule based. And this becomes important under certain situations like Upsell.

162
00:23:35.280 –> 00:23:36.470
Ranjit: So let’s say.

163
00:23:36.690 –> 00:23:59.770
Ranjit: Lonely planet has an agreement with Vfs, global Vfs, global, as you know, is a company that provides visa kinds of processing. And so any so so they you can set up a rule for lonely planet saying, anytime someone asked a question about visas. In that case I want you to forget about getting a response from

164
00:23:59.770 –> 00:24:11.589
Ranjit: From, you know, a model or something like that. But go route me to the particular service that Vfs global offers. And I can click on this service. And it will show me a video. And all of that.

165
00:24:13.240 –> 00:24:16.730
Ranjit: And the final question I’m going to ask here is a

166
00:24:18.710 –> 00:24:42.890
Ranjit: something that is kind of off. The wall has no bearing on travel, and The the goal here is to show you that we put put in place guardrails around scope. And so anytime there’s a question that’s considered out of scope relative to travel. The virtual assistant is able to detect that. And you know. Say, Hey, hold on, you know I’m a I. My focus is on travel.

167
00:24:42.930 –> 00:24:50.530
Ranjit: So that was the first demo I wanted to show you. Let’s go back and show you the now to the second demo. So

168
00:24:52.860 –> 00:24:55.119
Ranjit: so this was the travel assistant.

169
00:24:55.280 –> 00:25:01.539
Ranjit: The second demo I’m going to show you is a retail assistant. In this case the data source is structured.

170
00:25:01.650 –> 00:25:21.340
Ranjit: and it is stored in a postgres database in case you care about that. It’s it’s basically a retail database with dummy data. And and the objective of this chat Bot is to provide precise responses to any customer who comes and asks questions about the orders that they have placed.

171
00:25:21.370 –> 00:25:28.330
Ranjit: For this we have found that Gpt. 4 is a much better cost. You know, Roi

172
00:25:28.410 –> 00:25:33.800
Ranjit: model for you to use. And we’re so we’re able to use that model in in this particular scenario.

173
00:25:33.950 –> 00:25:50.010
Ranjit: the approach is basically, since you’re querying a structured database, you need the ability to take what the user says and automatically generate these queries that will run on the database and pull the pull the data for you, and then display that in a user friendly form into the chat board.

174
00:25:50.520 –> 00:25:59.410
Ranjit: and and so the highlight of this particular demo is that you can handle pretty complex database operations. And again, I’ll show you this in action

175
00:26:02.370 –> 00:26:06.329
Ranjit: and reset here, and then go back here and show you the

176
00:26:07.210 –> 00:26:09.800
Ranjit: what the backend looks like.

177
00:26:11.770 –> 00:26:15.050
Ranjit: So the first thing I wanted to do is just show you the database

178
00:26:15.230 –> 00:26:18.070
Ranjit: before I get into it. Let me start the retail assistant up

179
00:26:20.560 –> 00:26:25.560
Ranjit: and in, you know, for demo purposes we enable people to come in and

180
00:26:25.660 –> 00:26:32.399
Ranjit: you know, just randomly select some user. But let me give you kind of an insight into what this database looks like.

181
00:26:32.490 –> 00:26:34.750
Ranjit: The database has a

182
00:26:35.870 –> 00:26:43.790
Ranjit: 5 tables, customers. and we have about 37,000 customers who have placed about half a million orders

183
00:26:43.990 –> 00:26:50.369
Ranjit: which works out to about a million or items being ordered. There’s a product catalog that’s

184
00:26:50.620 –> 00:26:57.439
Ranjit: 2,500 or so and about 7,700 stores. So I’m going to come in and

185
00:26:57.790 –> 00:27:02.130
Ranjit: introduce myself as customer, Joan Kozak.

186
00:27:03.290 –> 00:27:05.990
Ranjit: So once I introduce myself as Joan Kozak.

187
00:27:06.140 –> 00:27:13.590
Ranjit: Meanwhile, on the left side of the screen, what I’m going to do is show you the information that’s available in the database for customer, Joan Cosi.

188
00:27:13.920 –> 00:27:15.819
Ranjit: So let me go in here

189
00:27:16.140 –> 00:27:18.279
Ranjit: and look at this, so

190
00:27:20.590 –> 00:27:39.269
Ranjit: you don’t have to pay attention to all this. I’m just doing this to show you that we’re using live data, and we’re not doing anything. No, no magic here. So here’s the information that’s in the back end database for customer, Joan Cozac. You’ll notice that she most recently placed an order on the seventh of November

191
00:27:39.810 –> 00:27:56.409
Ranjit: estimated ship date of the twelfth estimated delivery date of the eighteenth, and we have all the details of you know this order has 4 items in in different quantities. And each of these items has a way to, you know, price and a color and and a category and all that stuff.

192
00:27:56.810 –> 00:28:12.219
Ranjit: So let’s come back here the first point I wanna make is that you know it’s it’s a bit of a nuance. But you know we have the ability to recognize and address someone if if we can recognize their gender. So in this case, Joan was interpreted to be a missus.

193
00:28:12.360 –> 00:28:25.909
Ranjit: the next. So the first question that I so I, Joan, am asking is, When will my most recent order be? Ship? So here, like, I explained previously. We take it, we take the question.

194
00:28:25.940 –> 00:28:39.049
Ranjit: and then behind the scenes this gets converted into into kind of a quasi SQL. Query and an appropriate Prom, and then, which constructs a prompt that gets sent back to the GPT. 4 model on the back end.

195
00:28:39.350 –> 00:28:43.760
Ranjit: So you’ll notice that the most recent order was placed on the seventh.

196
00:28:44.310 –> 00:28:50.380
Ranjit: and is expected to be expected to be delivered on the twelfth. Sorry! Expected to be shipped on the twelfth.

197
00:28:51.550 –> 00:28:55.739
Ranjit: The next question, Joan asks, is, when will it be delivered?

198
00:28:57.200 –> 00:29:10.009
Ranjit: And again, you’ll you’ll notice now that the it’s it’s all in the context of this order I didn’t have to repeat. When will my most recent order be delivered and things like that? And you’ll notice that the date is the eighteenth, and and that’s what it shows.

199
00:29:10.130 –> 00:29:11.109
Ranjit: So now

200
00:29:14.230 –> 00:29:18.319
Ranjit: Joan says. Oh, no, I have to attend the wedding on that day.

201
00:29:18.460 –> 00:29:21.790
Ranjit: Now, in a normal chatboard the chat port

202
00:29:22.000 –> 00:29:28.780
Ranjit: would respond and say, What are you talking about? We were just talking about products, but in a conversational type situation.

203
00:29:29.050 –> 00:29:36.010
Ranjit: When you ask about when something is going to be delivered. and then you follow it up by saying you’re you’re attending a wedding on that day.

204
00:29:36.120 –> 00:30:04.039
Ranjit: The implication is that you won’t be available on that day to pick up your order, and this kind of inference and reasoning is supported by our conversational engine. What this does is based on this, we give the user the opportunity to reschedule the delivery, or, you know, have it delivered to for pickup or something like that. So I just wanted to show you this. I’m not going to connect to a live agent yet at this point, because we showed you that in the previous so

205
00:30:04.690 –> 00:30:05.800
Ranjit: example.

206
00:30:07.920 –> 00:30:14.249
Ranjit: So the next question I have is, what is in the order? And coming back to the database on the left side of my screen.

207
00:30:14.440 –> 00:30:17.690
Ranjit: you’ll notice that there’s like 4 items.

208
00:30:17.720 –> 00:30:29.350
Ranjit: And there’s like a camera and a skin case and all of these kinds of things. And you can see that the a Chatbot was able to actually take this very simple question

209
00:30:29.540 –> 00:30:41.209
Ranjit: and convert it into a somewhat complicated query on the back end that said, what were the items in the order placed by Joan Kozak in her most recent order. So so there’s a lot of heavy lifting going on in the backend.

210
00:30:42.410 –> 00:30:43.480
Ranjit: So now

211
00:30:43.820 –> 00:30:47.099
Ranjit: Joan says, how much did I pay?

212
00:30:47.140 –> 00:30:50.860
Ranjit: Now, this information is not in the database. It has to be computed.

213
00:30:50.980 –> 00:30:59.639
Ranjit: So we have 4 items each of those. So these are the 4 items. Each of these these items has a quantity which is 1, 2, 2, and 7

214
00:31:00.120 –> 00:31:12.590
Ranjit: and a price. So the system has to actually do some computation on the back end. Of multiplication and addition to come up with a total for the order. And and, as you can see, it’s capable of doing that as well.

215
00:31:12.990 –> 00:31:17.260
Ranjit: So let me ask another question.

216
00:31:18.610 –> 00:31:25.079
Ranjit: Is it possible for me to pick up the order from somewhere instead? And in this case we will.

217
00:31:25.280 –> 00:31:35.260
Ranjit: you know the the system has an understanding of the customer’s location and based on the location they’re they’re suggesting the the store that’s most

218
00:31:35.420 –> 00:31:36.670
close by.

219
00:31:36.960 –> 00:31:38.970
Ranjit: So let me, just.

220
00:31:39.720 –> 00:31:44.180
Ranjit: you know, go go and have, you know, more conversational experience. Here.

221
00:31:44.220 –> 00:31:47.510
Ranjit: let me see, I’m just trying to find.

222
00:31:48.670 –> 00:31:49.500
Ranjit: Okay.

223
00:31:52.130 –> 00:31:54.700
Ranjit: what categories of products have I purchased?

224
00:31:54.770 –> 00:32:00.659
Ranjit: Now this is you’ll you’ll notice that there’s a bunch of categories here. And

225
00:32:00.970 –> 00:32:13.480
Ranjit: and and you know a. This is a fairly complex query where you have to do some grouping and some, you know, getting out unique values and all of that. So these kind of complex queries are also possible on the back end.

226
00:32:13.640 –> 00:32:22.390
Ranjit: I’m going to say. what categories of products have I purchased which are not appliances or devices.

227
00:32:22.920 –> 00:32:30.000
Ranjit: So if you look at this list, you’ll notice that the only ones that kind of qualify for that are movies and and games.

228
00:32:30.200 –> 00:32:43.340
Ranjit: I’m gonna try asking another question. II don’t know if there’s an answer for this. But just to show, you know that I can ask pretty complicated questions, what categories of products I purchase, which are not appliances or devices, but which are pink

229
00:32:43.740 –> 00:32:50.570
Ranjit: and so it’s able to tell me that I have purchased the following pink product. And if you go into this database.

230
00:32:50.710 –> 00:32:53.540
Ranjit: so this is a pink. But this is a cell phone.

231
00:32:53.680 –> 00:33:03.420
Ranjit: There’s a an another pink, which is smartphone. This is the only pink item which is a game and not a a device or an appliance.

232
00:33:04.060 –> 00:33:09.729
Ranjit: Finally, I wanted to show you the ability to to have guardrails on the data. So

233
00:33:09.780 –> 00:33:14.610
Ranjit: maybe I let’s say, I say something like. what items have my neighbors bought?

234
00:33:15.390 –> 00:33:23.249
Ranjit: The system says is able to detect that look. II can only give you information about your order and not about any other stuff.

235
00:33:23.290 –> 00:33:32.589
Ranjit: And then, you know, I can ask kind of a completely off the wall out of context question, and because of the scope guardrails that are in place

236
00:33:32.650 –> 00:33:37.939
Ranjit: the system should reject that and just say, Hey, you know, just ask me retail questions.

237
00:33:39.310 –> 00:33:40.040
Ranjit: Okay.

238
00:33:41.490 –> 00:33:44.459
Ranjit: so okay, so we’ll leave that for now

239
00:33:44.510 –> 00:33:47.789
Ranjit: coming back to the demo

240
00:33:48.420 –> 00:33:49.290
Ranjit: up.

241
00:33:50.350 –> 00:33:51.120
Yeah.

242
00:33:51.330 –> 00:33:53.629
Ranjit: I wanted to jump into the next demo.

243
00:33:54.620 –> 00:34:01.619
Renga: Here we go, and we already have some questions coming in ranges. But I’m going to park that after the end of your demos

244
00:34:01.640 –> 00:34:06.910
Ranjit: got it? Got it? Yeah. So I have a I have. How much longer do I have?

245
00:34:06.970 –> 00:34:08.710
Ranjit: Just can you tell me the time.

246
00:34:09.020 –> 00:34:15.579
Renga: Yeah, you have 30 more minutes London. I suspect we might want to leave 10 to 15 min for Christians.

247
00:34:15.630 –> 00:34:20.460
Ranjit: No problem, no problem. Yeah. So I think II think we, we should be good with this.

248
00:34:21.580 –> 00:34:28.049
Ranjit: So let me talk about the third demo. The third demo is to kind of do the same thing we did.

249
00:34:28.280 –> 00:34:47.630
Ranjit: We first showed you unstructured data, then showed you structured data. But this demo is to look at Pdf documents there. There are some important use cases here. We have, you know, banking customers who who need to consult various Pdf documents anytime. Say, for example, there

250
00:34:47.630 –> 00:35:05.640
Ranjit: compiling documents for a loan and things of that sort. There’s many procure to pay types of use cases that use Pdf documents. There’s even situations where a customer might find value in looking up a bunch of Pdf manuals. So.

251
00:35:05.780 –> 00:35:19.820
Ranjit: having the ability to look at unstructured Pdf documents is very useful in this case. We’ve decided to use the Google Palm 2 model. And Google has really, you know. So in in our evaluation

252
00:35:20.130 –> 00:35:30.950
Ranjit: of all of these different products, Google did the best job when it came to the ability to extract information from Pdf files, including follow-up questions.

253
00:35:31.490 –> 00:35:32.580
Ranjit: And

254
00:35:32.630 –> 00:35:47.950
Ranjit: the approach we have is relatively more traditional in the sense that what we’re doing is vectorizing the information that’s in those Pdf files and then using our middleware to to have an in to enable the interaction between the customer and and the Pdf.

255
00:35:48.170 –> 00:35:57.419
Ranjit: And and the other benefit of using a model like Palmtwo is that it’s 5% as expensive as GPT. 4, with equivalent quality, and in some cases better

256
00:35:57.550 –> 00:36:01.740
Ranjit: so let me show you how that works. It’s the demos a little bit different.

257
00:36:03.910 –> 00:36:13.600
Ranjit: The the ui looks a little bit different is what I mean to say. So let me come in here, and this is just a really straightforward

258
00:36:14.520 –> 00:36:15.210
Ranjit: you know.

259
00:36:16.110 –> 00:36:21.389
Ranjit: Demo. So let me maybe reload the page or something like that. Just so it’s clean.

260
00:36:21.780 –> 00:36:27.289
Ranjit: I’m going to ask a new question. And I’m just going to copy paste these questions from

261
00:36:30.580 –> 00:36:39.869
Ranjit: from another document. So this is we have loaded up a document which represents the Hyundai Ionic

262
00:36:39.920 –> 00:36:51.000
Ranjit: Vehicle Manual. So there’s a collection of vehicle manuals that we’ve indexed and people can ask questions about that. So the first question is, what does the evasive steering assistant do? And

263
00:36:51.130 –> 00:36:57.349
Ranjit: while I’m doing? While I do this, let me also show you what what the actual document looks like. This is the

264
00:36:57.400 –> 00:37:02.140
Ranjit: actual document. And if you go to page 3, 63,

265
00:37:02.290 –> 00:37:08.900
Ranjit: it gives you information on what the evasive steering assist function does. So you’ll notice that it says.

266
00:37:09.010 –> 00:37:16.709
Ranjit: You know, when a risk of collision is detected, it will want the driver and all of that I can. I can sort of ask a follow up question which is

267
00:37:16.820 –> 00:37:18.350
Ranjit: like.

268
00:37:18.910 –> 00:37:20.780
Ranjit: does it work for pedestrians

269
00:37:22.140 –> 00:37:31.949
Ranjit: and and and you’ll you’ll notice that it’s able to understand that in the context of what I asked about the evasive steering assistant, it can actually work for pedestrians as well.

270
00:37:33.400 –> 00:37:36.120
Ranjit: Let me ask a couple more questions.

271
00:37:37.220 –> 00:37:40.330
Ranjit: So what adjustments can I make on the front headrests?

272
00:37:44.430 –> 00:37:50.399
Ranjit: So in this case? There’s some information on page 117 about headrest and things like that.

273
00:37:50.420 –> 00:38:09.940
Ranjit: So it says, you can adjust, you know. Adjust them by raising or lowering them and moving them forward, backward. And, by the way, in this case, I’m just giving you text text responses. But if there are relevant images, we can also present those images back in the response. So you, you know you get a good sort of multimedia effect.

274
00:38:10.100 –> 00:38:14.149
Ranjit: of of just simply searching through various Pdf documents.

275
00:38:14.720 –> 00:38:16.389
Ranjit: The next thing I ask is.

276
00:38:18.080 –> 00:38:23.189
Ranjit: okay. Let me ask a follow up. Can I recline the seat back towards the front?

277
00:38:24.310 –> 00:38:34.890
Ranjit: Now, what I’m trying to get at is, you know. If if my headrest is too high. If if you can imagine a car, let’s say I’m on the front, and and I raise my headrest too high.

278
00:38:35.210 –> 00:38:41.999
Ranjit: and I reply, and I move it in the front. It has the. It might hit the windshield. I’m kind of looking

279
00:38:42.130 –> 00:38:49.410
Ranjit: looking to get that response. So let let me ask a follow up question. What happens if I do this with the headrest raised.

280
00:38:50.180 –> 00:39:02.510
Ranjit: and then here, finally, you know II get the response that I’m looking for. If I if it, if I recline it towards the front with the seat cushion, raise the the headrest will come in contact with the

281
00:39:03.450 –> 00:39:09.879
Ranjit: with the visor, as as it as shown over here again, this is just a an example of

282
00:39:10.170 –> 00:39:32.830
Ranjit: where we can take data from all sorts of different documents. And by the way, you can mix and match all of these, you, you may have some structured data, some unstructured data and some Pdf data, and you can mix all of them into one single chat pot that then goes and pulls the appropriate information. In addition, we can look at data from salesforce service. Now. Jira, all of these different

283
00:39:32.830 –> 00:39:40.360
Ranjit: back-end systems as well as extract text from videos assuming that, you know, there’s.

284
00:39:40.520 –> 00:39:45.300
Ranjit: you know, there, there’s people speaking, and and some kind of text in the video.

285
00:39:46.040 –> 00:39:48.960
Ranjit: Alright. So that was my third demo.

286
00:39:50.230 –> 00:39:55.340
Ranjit: I have a couple more that are going to be really quick. The next demo I wanted to show you

287
00:39:55.450 –> 00:40:01.939
Ranjit: was the ability to do field extraction. And what this really means is that I can send a

288
00:40:02.400 –> 00:40:06.329
Ranjit: a sort of semi-structured Pdf document

289
00:40:06.360 –> 00:40:07.849
Ranjit: into my system.

290
00:40:07.950 –> 00:40:33.029
Ranjit: and I can automatically extract some important fields. And so let me set this up for you. okay, before I do that. So I’ll explain what what we, what we do here, we’re using the Google Palm 2 model. Again, this does an excellent job of this kind of field extraction. In addition, we’re using kind of a toolkit that is available as part of our platform to do language processing and the ability to to manage workflows.

291
00:40:33.580 –> 00:40:41.479
Ranjit: In this case there’s also no training required. So, in other words, I can just specify what fields I want extracted.

292
00:40:41.700 –> 00:40:49.920
Ranjit: Throw a document at at the platform, and it will extract and display the fields for me. So let me show you that with a demo

293
00:40:50.390 –> 00:40:55.659
Ranjit: in this case, I’m going to do something a little bit different. I’m going to be a user.

294
00:40:56.730 –> 00:40:58.330
Ranjit: And I’m gonna be sending

295
00:40:59.760 –> 00:41:02.470
Ranjit: to my email gateway.

296
00:41:04.370 –> 00:41:06.119
Ranjit: A. US. Bank statement.

297
00:41:07.240 –> 00:41:10.410
Ranjit: And I’ll show you this. Statement in a second.

298
00:41:11.200 –> 00:41:12.700
Ranjit: I’m going to attach this

299
00:41:17.540 –> 00:41:20.419
Ranjit: and get this ready and send it. So here’s here’s what’s gonna happen.

300
00:41:20.600 –> 00:41:41.879
Ranjit: This document is going to go to the back end, and at the back end you have someone who will inspect the fields that have been extracted from this document in order to make any decision. Maybe they want to make a decision on loans, or you know, or something about opening an account, or whatever it is. But the important point is that we’re taking an unstructured Pdf document.

301
00:41:41.980 –> 00:41:48.889
Ranjit: And you know, automatically extracting the fields. And we’re doing this through email. So I’m gonna hit, send on this email.

302
00:41:49.970 –> 00:42:01.159
Ranjit: And coming back here, let me just kind of show you this email gateway that we have. This email gateway should start showing some activity here at the top as this new email comes into it.

303
00:42:02.180 –> 00:42:02.930
Ranjit: And

304
00:42:03.310 –> 00:42:07.929
Ranjit: so, as you can see it, it just woke up. There’s Usbank statement.

305
00:42:08.270 –> 00:42:14.110
Ranjit: And so this you know, text is sent to the back end.

306
00:42:14.170 –> 00:42:22.279
Ranjit: and if I click on this, you know, you should see that you know all. I’m getting greens, which means everything kind of worked out fine.

307
00:42:22.320 –> 00:42:24.739
Ranjit: I’m gonna log in in the back end

308
00:42:26.330 –> 00:42:30.680
Ranjit: as a user who? Whose job it is to look at these incoming documents.

309
00:42:31.260 –> 00:42:41.889
Ranjit: So let me just click on all the tasks. You’ll see a new field extractor task got created for me a few seconds ago. So this is a situation where an incoming email can create

310
00:42:42.010 –> 00:42:46.380
Ranjit: a task for a a you for for one of your employees.

311
00:42:47.010 –> 00:42:49.050
Ranjit: I click on this.

312
00:42:51.080 –> 00:43:12.740
Ranjit: and you’ll notice that it shows a bunch of different fields here. So let me just open this, and you’ll see that it gives me a bank address and a bank phone number and the customer address, and you know, account number and all of these kinds of things. I can also visually look at this document, and and this is this is a document that was mailed. So this is semi-structured document.

313
00:43:12.760 –> 00:43:17.829
Ranjit: But all I had to do was specify which fields I want extracted.

314
00:43:18.130 –> 00:43:26.970
Ranjit: emailed this document in and it automatically extracted those fields and put them into into this form that I show you. So, for example, if I look at the ending balance.

315
00:43:27.220 –> 00:43:30.069
Ranjit: I look at ending balance 10,500.

316
00:43:30.630 –> 00:43:37.149
Ranjit: So that’s 10,500. It can even do some things like counts. So if I look, you know the the number of checks that were paid.

317
00:43:38.470 –> 00:43:40.340
Ranjit: Is 3,

318
00:43:40.410 –> 00:43:50.770
Ranjit: the largest transaction, for example, 3,600 bucks, and you’ll see that of all the transactions that is the largest one, and in addition, you know, the ability to extract

319
00:43:51.110 –> 00:43:53.970
Ranjit: customer address 1, 2, 3, 4, anywhere. Drive

320
00:43:54.260 –> 00:43:59.160
Ranjit: should be somewhere here. Yeah, there it is. And and so

321
00:43:59.300 –> 00:44:22.669
Ranjit: so the system is able to pull all of this information out, which is a huge time saver this can be used in very much in a procure to pay kind of situation where you want to initiate the payment cycle by emailing an invoice. In this case, I showed you a bank statement. You can initiate. banking transactions by mailing bank statements. We can do invoices. We can do many other kinds of these

322
00:44:22.670 –> 00:44:36.920
Ranjit: semi-structured documents. All we have to do is define the fields that need to get extracted, and this can be done without training. So, in other words, I don’t need to ask you, for, you know, give me a hundred sample documents so that I can figure out what the patterns are and things like that.

323
00:44:37.990 –> 00:44:42.400
Ranjit: Alright. Coming up on the final demo, which is really going to be very quick.

324
00:44:42.770 –> 00:44:59.910
Ranjit: okay. The final demo I wanted to show you is something that is unique to autonomy. We have the ability for you to express workflows. Like, I said, we have a workflow platform. What this does is allow you to create workflows.

325
00:45:00.100 –> 00:45:03.220
Ranjit: deploy them, run them for your customers and all of that.

326
00:45:03.370 –> 00:45:22.539
Ranjit: We now have the ability for you to actually speak into a, or type into A into a prompt box and based on the information that prompt box, we automatically generate the work flow for you, which which is a huge time saving capability. And it’s also a unique capability in this industry. If you look at the entire

327
00:45:22.540 –> 00:45:38.679
Ranjit: Enterprise, workflow automation industry, we are the first ones to do this. And again, we’re using the model from Google, which which now, you know, the chat bison model has a a huge token limit. And all of that. It’s it’s really great.

328
00:45:38.770 –> 00:45:39.590
Ranjit: So

329
00:45:39.690 –> 00:45:53.589
Ranjit: to be a little technical. What this does is generate a syntax called Bpm under the hood, and this Bpm. And is a standard notation that can be loaded into your into your tool. So I’m going to just show you that really, quickly.

330
00:45:53.650 –> 00:45:57.790
Ranjit: And this will be my final demo. So let’s go back here

331
00:45:58.300 –> 00:46:00.070
Ranjit: and option

332
00:46:01.530 –> 00:46:02.510
Ranjit: open up

333
00:46:02.760 –> 00:46:16.160
Ranjit: this. So this is our. It’s sort of an empty canvas. And into this canvas I can actually type in the prompt that I want. So let me just give me a second here, I’m gonna copy paste. I’m gonna start with something very simple.

334
00:46:16.310 –> 00:46:24.830
Ranjit: I’m just gonna say, create a Bpm and workflow for a travel booking process. So this is not technically customer support. But I just wanted to show you this as an example.

335
00:46:26.030 –> 00:46:31.649
Ranjit: it says, process should include steps for selecting destination, choosing the dates and making payment.

336
00:46:32.460 –> 00:46:34.699
Ranjit: So when I submit this

337
00:46:34.840 –> 00:46:38.830
Ranjit: behind the scenes. what we do is we take this prompt

338
00:46:39.030 –> 00:46:56.739
Ranjit: and do a lot of different things behind the scenes to actually create a workflow. So now you’ll notice that this is automatically created a workflow for you and the the purple blocks in the workflow are pieces where users have to actually provide input

339
00:46:57.240 –> 00:47:07.710
Ranjit: and the yellow blocks are those where you know you have some third party. You can say it calls an Api or a service, and the system has actually been able to, you know.

340
00:47:07.920 –> 00:47:10.210
Ranjit: make a distinction between.

341
00:47:10.480 –> 00:47:17.129
Ranjit: you know the types of blocks that are required, and and show you the appropriate one. Let me show you one that’s a little bit

342
00:47:17.180 –> 00:47:29.739
Ranjit: more complicated and and more in the context, you know, in the context of customer support generally. Yeah, you know, let’s say, a bug report comes in. Maybe you have service now, salesforce and an sap, and all of these kinds of tools

343
00:47:29.750 –> 00:47:43.349
Ranjit: what you can do is create a work flow that looks exactly like this. And you know, and and our platform provides you with all the adapters that you need to connect to all these third party platforms as well. So I’m gonna clear the screen here

344
00:47:44.700 –> 00:47:50.210
Ranjit: and type in now the workflow that I just showed you

345
00:47:50.960 –> 00:47:59.799
Renga: in terms of, would you mind just zooming in a bit so that we know once the output comes through?

346
00:48:00.170 –> 00:48:01.130
Ranjit: Yeah.

347
00:48:07.420 –> 00:48:08.440
Ranjit: So here

348
00:48:08.810 –> 00:48:11.539
Ranjit: you’ll notice that the system.

349
00:48:12.540 –> 00:48:19.669
Ranjit: you know. Here’s what I said. When an email is received, do the following steps respond immediately acknowledging receipt. Does that

350
00:48:19.750 –> 00:48:21.429
Ranjit: if this is a bug report.

351
00:48:21.520 –> 00:48:25.799
Ranjit: create a service. Now take ticket and update salesforce. It does that

352
00:48:26.050 –> 00:48:30.480
Ranjit: if and if the attachment is an invoice, create a new invoice in SAP.

353
00:48:30.840 –> 00:48:40.000
Ranjit: Wait for 10 min and then create a new task in autonomate. Now, it’s it’s routed it a little differently than a human might have. But it’s it’s

354
00:48:40.260 –> 00:48:56.200
Ranjit: remember to put a 10 min timer to wait. It’s remember to put in these kinds of email recei receiving gateways. It’s it’s created these business logic branches. And it’s created all the tasks that are required in this workflow. So this is.

355
00:48:56.260 –> 00:49:11.130
Ranjit: yeah, a quite a remarkable innovation when it comes to the speed with which you can create and deploy workflows. So I just wanted to show this to you, because this is something that we’ve just recently rolled out. It’s in production, and if you use our platform you’ll be able to check this out.

356
00:49:11.580 –> 00:49:17.729
Ranjit: That’s all I had from a demo perspective. How are we doing on time, Ranga? I had a couple of more slides.

357
00:49:18.520 –> 00:49:24.600
Renga: We have 15 min, do you just please through them quickly, so that we have room for questions.

358
00:49:24.690 –> 00:49:36.190
Renga: Yeah, sorry. Just 2 min. Because this is I. This was mentioned in the agenda, and I just wanted to make sure I covered it. So let me just talk.

359
00:49:36.210 –> 00:49:39.140
Ranjit: and I’ll finish this in 2 min, I promise.

360
00:49:39.310 –> 00:49:48.520
Ranjit: So the first so I want to make a definition which is borrowed from the world of the stock market in the stock market. There’s a concept called efficient frontier.

361
00:49:48.690 –> 00:49:54.450
Ranjit: What this means is that in order to maximize your reward

362
00:49:54.560 –> 00:50:01.770
Ranjit: and reduce your risk. You set certain parameters, and then the system automatically buys and sells stocks for you.

363
00:50:01.810 –> 00:50:04.370
Ranjit: based on that efficient frontier.

364
00:50:04.560 –> 00:50:11.900
Ranjit: To to accommodate your risk reward specifications we want. We are applying that same approach.

365
00:50:12.230 –> 00:50:20.150
Ranjit: Or when you’re using models, we have access to dozens of different models through our Llm. Console

366
00:50:20.260 –> 00:50:29.840
Ranjit: and based on your you as a customer based on your requirements in terms of data security in terms, in terms of cost, in terms of quality. In terms of response speed.

367
00:50:30.020 –> 00:50:33.600
Ranjit: we automatically select the appropriate model

368
00:50:33.800 –> 00:50:45.849
Ranjit: that will work for you. Now you can say, I only want to use. GPT. 4. And that’s fine. But for customers who are willing to or need that kind of flexibility, we provide this capability in our tool.

369
00:50:46.280 –> 00:50:47.640
Ranjit: I’m just gonna

370
00:50:48.110 –> 00:51:01.969
Ranjit: I’m not gonna say much about these slides. I’m gonna stick them up here. But we have a comprehensive approach to supporting security in terms of data redaction in terms of sending, you know, sending limited data to third party Llms,

371
00:51:02.030 –> 00:51:21.119
Ranjit: and we have a like, I said, an approach, a comprehensive approach to cost maintenance as well. And the reason I’m putting up these slides here is because so that they show up in the recording, and for for all of you. You you’ll have. You’ll get a copy of this and be able to go through these at leisure and ask us questions.

372
00:51:22.010 –> 00:51:34.000
Ranjit: this is a a chart that shows you. Ha! You know, kind of an overview of all the different Llms that we support. Okay, I will stop now, Ranga, and happy to answer questions.

373
00:51:34.390 –> 00:51:56.640
Renga: Actually, would you mind keeping the Llm selection considerations? Full screen? Because, I’m I’m just going backwards first. So tackling the question which is freshest at this point. Offense to the questions lined up before that. But it seems to be repert to Llm. Selection, too. So the first question that the audience had for you

374
00:51:56.760 –> 00:52:14.969
Renga: is, I believe, our audiences feeling shy. So they have all asked it anonymously. But the first question that came in is, can 8 make new custom gpt models to work with my tech stack as per the announcement from Openai dev day last week.

375
00:52:16.090 –> 00:52:34.630
Ranjit: Yeah. So we we actually, before, before the announcement came out, we had our own technique for creating these custom gpts. We on the back end, we have some configuration where we can. You know, selectively, pick the data and then, use that to.

376
00:52:34.850 –> 00:52:42.029
Ranjit: II think the word is ground the GPT. Model or the the Llm. To ground the model so that

377
00:52:42.410 –> 00:52:54.799
Ranjit: it responds to those questions we have looked at the announcements the recent announcements from Openai, and we believe that that makes our job a little bit easier. But we haven’t yet incorporated that into our application.

378
00:52:56.240 –> 00:53:23.009
Renga: Sure, and I believe we would probably also want to include an update this slide for Gbt turpo, I believe. So that the pricing is included for those 2. That’s that’s that’s a good point, I think. You know, primarily due to, you know, column restrictions on this slide. I kind of summarized everything and everything into a single column. But there’s multiple columns like I said, we have dozens of

379
00:53:23.080 –> 00:53:42.960
Ranjit: different models that we can invoke including llama falcon. Yeah, you know all the different hugging face models and things like that. I just picked a few the the goal here is actually more to look at the criteria by which we evaluate them than it is to look at any specific

380
00:53:43.290 –> 00:54:00.119
Ranjit: a specific model. But the reality is that it’s much easier to use these third party models. As if you wanna set up your own. You know, if you if you have a sensitivities around security, for example, and you want to set up your own models privately. That’s doable. But that’s expensive.

381
00:54:00.870 –> 00:54:16.060
Renga: right? And actually, that’s a good segue into the next question that our audience had, and they wanted to know what is our pricing model? In, you know, within, in the landscape of also talking about the cost to be able to consume these Lms, too.

382
00:54:16.170 –> 00:54:24.930
Ranjit: Right? So if you are. If you are using only the third party elements like Gpt. 354 palm

383
00:54:25.330 –> 00:54:36.010
Ranjit: and things like that. The pricing model is basically the the cost of the transaction with a with a commission on top of it. It’s a transactional model. And what we do is

384
00:54:36.060 –> 00:54:39.180
Ranjit: you know. let’s say

385
00:54:39.200 –> 00:55:05.149
Ranjit: you run a website, and you you you know your customers ask 100,000 questions in a month, and half of them go to Gpt. 3.5 and half of them go to GPT. 4, for example, what we would do is just charge you the cost for that with a little commission. The reality is that when you actually use it. We don’t go back to the Llm’s. For every question that you ask. We do a lot of what we call response caching.

386
00:55:06.000 –> 00:55:08.809
Ranjit: Let me show you that

387
00:55:09.230 –> 00:55:10.770
Ranjit: that allows you

388
00:55:10.960 –> 00:55:29.929
Ranjit: to preserve results from Llms so that you don’t have to go back to it again. So let’s say you have, you know, nor the normal approach. When when you know for any website, for any customer any website where customers interact, you have probably 50 to 100 questions that are kind of the the bulk. And then you have this long tail.

389
00:55:30.010 –> 00:55:45.680
Ranjit: So what we do is we we help you create and cache responses for these top 50 or 100 questions. So you don’t even need to go back to Lm, or you can say, Okay, I’ll only refresh these questions every 5 days and things of that sort. So

390
00:55:45.800 –> 00:55:59.340
Ranjit: so the actual cost, even if you have a hundred 1,000 queries, would would probably work out to the cost of maybe 5 to 10,000 queries based on our caching capability. And that sorry, long answer hopefully, that was useful.

391
00:56:00.040 –> 00:56:11.400
Renga: No, that’s interesting. And that also forms the next segway here into the next question, too, is that 1 one person is, how quickly can you set up a custom chat board

392
00:56:11.580 –> 00:56:12.949
Renga: for our website.

393
00:56:13.930 –> 00:56:31.639
Ranjit: so as I mentioned previously, so if if it’s website data and it’s public data, since we don’t we don’t scrape your site, or do that. We only look up your site when when a question gets asked so we configure this configuring. Our new site

394
00:56:31.640 –> 00:56:46.699
Ranjit: is really, honestly, an hour’s work. So all we have to do is you give us a list of websites and pages that you want to have indexed, we add it to our index, and you can have a chat bot within an hour that supports that particular. You know

395
00:56:46.760 –> 00:56:51.249
Ranjit: customize data. If you’re if you’re dealing with structured data

396
00:56:51.450 –> 00:56:59.300
Ranjit: building a, you know, creating a chat board for that that can take a little bit longer, because we need to understand and map

397
00:56:59.360 –> 00:57:10.739
Ranjit: the schema or the data dictionary of your structured data source and that can take some time it. It takes more time if you’re if you if if you have a custom database. But if you’re using something like

398
00:57:10.810 –> 00:57:16.580
Ranjit: service now, or salesforce, it’s much quicker, because we are we. We understand those schemas

399
00:57:16.690 –> 00:57:32.510
Ranjit: right away. And then for Pdf, if there’s a large number of Pdf files, it just takes time to index these files and create, you know, the vector database for for these files. So again, we’re talking about, you know, a few days to a week. Kind of approach.

400
00:57:35.320 –> 00:57:53.929
Renga: Thank you, I, Ranjith, and I believe we answered most of the questions which came in the Q. And a panel. We still have 5 min, so I do have folks in the audience. Anybody who wants to answer or ask a question live to Runjit. I can selectively enable you so that he can speak up and ask him yourself

401
00:58:00.690 –> 00:58:04.090
Renga: going ones going twice

402
00:58:08.490 –> 00:58:18.750
Renga: alright in case you have any questions, and you’ve probably not had the chance to ask it. Live, we’ll still be circulating the recording and snippets of of this particular webinar.

403
00:58:18.800 –> 00:58:28.329
Renga: for who has been able to register for these sessions, and you have a chance to just respond to those emails and also come ask us questions over email.

404
00:58:28.500 –> 00:58:37.919
Renga: But if we don’t have any questions, I’m just going to pause for 5 more seconds. If somebody wants to ask a question, either through the QA. Widget, or

405
00:58:38.050 –> 00:58:43.500
Renga: also I’m just live. I’m going to give 5 s here.

406
00:58:48.520 –> 00:59:09.329
Renga: right? Looks like it’s a it’s a slow Thursday, Ranjit. But let folks come to us in due time. But thank you so much for taking us through this session and patiently asking all and answering all these questions for the folks who are listening in, we will definitely send over the recording and give you a contact info, so that you are able to.

407
00:59:09.360 –> 00:59:18.940
Renga: you know. Come back to us with any questions that you might want to ask later. Thank you, and thank you for joining us. Have a wonderful evening for your time. Thanks, Ranga.

408
00:59:19.680 –> 00:59:23.229
Renga: Thank you. Take care everyone bye, bye, engine.