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

Harnessing Hyper-automation and Gen AI for Lasting Customer Relationships

  • Date

    25 Oct'23
  • Time

    11:00 AM - 12:00 PM IST
  • Duration

    60 mins

00:12:11.940 –> 00:12:26.330
Ranjit: okay? So so let me get into Pop in the presentation mode. So I’ll start by describing what we do at auto autonomate. Our slogan is hyper automating. The custom, you know, is to enable the autonomous enterprise.

00:12:26.330 –> 00:12:46.539
Ranjit: And the idea here is that in very much the same way that you saw the evolution of vehicles to becoming autonomous where they they had intelligence, the ability to make decisions, the ability for you know, man and machine to function together. And all of that, we want to bring many of those same principles into enterprises. So we’ve called that the autonomous enterprise

00:12:46.700 –> 00:12:56.450
Ranjit: and like, I like Ranga introduced today, the focus is on. You know how we can use genii to hyper automate customer support.

00:12:57.620 –> 00:13:06.489
Ranjit: So let me take a step back and describe kind of it at a high level. You know, the architecture of the autonomate platform.

00:13:06.520 –> 00:13:27.550
Ranjit: And we, we’ll focus on one piece of it today, which is kind of the front end part of it. But if if you take a step back and and look at workflows in general, if you look at any customer journey, any any workflow that is deployed in service of a customer’s requirements. There’s broadly 3 3 important pieces in that

00:13:27.880 –> 00:13:33.109
Ranjit: on the front end. You have what we call the intake piece where a customer comes in, and

00:13:33.180 –> 00:13:51.200
Ranjit: communicates to the enterprise what they want. This could be something as simple as a custom, you know, as a a you know it. Help desk request, or it could be something as complicated as applying for a loan. So there’s a wide variety of you know these customer journeys that the intake process needs to support.

00:13:51.200 –> 00:14:15.180
Ranjit: And in our platform we give, we, we provide the tools for you as an enterprise to create these these intake mechanisms. These can be through many different channels, through many different modes. And you know, so as as indicated here. You can connect through, you know, chat. And you can connect through text and SMS web, mobile voice text all of those kinds of things.

00:14:15.550 –> 00:14:26.740
Ranjit: So. Yeah, on the intake side you you can have a kind kind of either a what we call a virtual channel through like a chat board or a live channel like a contact center.

00:14:26.970 –> 00:14:31.520
Ranjit: Now, once the information from the customer has been received by the back end.

00:14:31.640 –> 00:14:44.730
Ranjit: we have a powerful business process management engine that is used to fulfill this customer’s request. So this is kind of the workhouse of our platform. It’s back end work, flow, enterprise, work, flow automation.

00:14:44.920 –> 00:15:01.909
Ranjit: And you know, we provide a a really powerful capability that uses Vpm kind of as the underlying technology and gives you the ability to create all of these customer journeys on the back end. So, for example, if someone’s applying for a loan

00:15:02.210 –> 00:15:25.420
Ranjit: on the front end, they’ll use the intake mechanism to give you the information, and on the back end you can put together all the orchestration that’s required. All the you know credit officer and sales officer and appraiser, and all of these different people, all of these people with different roles, can come together to fulfill this particular workflow. So we give you all the tools to create this using kind of a low code paradigm where

00:15:25.520 –> 00:15:34.259
Ranjit: you can drag and drop different boxes in in our a studio application to to to build and deploy these workflows.

00:15:34.880 –> 00:15:47.390
Ranjit: and on the back end, when you have these workflows, they generally integrate to the whole bunch of different tools like CRM. Ticketing, banking, and so on. Once your workflow or your customer journey is deployed.

00:15:47.660 –> 00:15:53.049
Ranjit: we also provide a real-time analytics infrastructure that allows you to focus, you know, to

00:15:53.080 –> 00:15:58.209
Ranjit: focus on different pieces of the customer’s journey. Figure out where they’re having success.

00:15:58.410 –> 00:16:09.099
Ranjit: where they are not. And you know, and how you can continuously improve this customer journey over time. So this is kind of the the big picture of what our platform is.

00:16:10.110 –> 00:16:28.810
Ranjit: so let me shift gears now to talking about how generative AI plays in, you know, as part of our platform for for this. You know, for this. Webinar, I’m going to focus a lot on the intake side of the of of this

00:16:28.850 –> 00:16:34.830
Ranjit: architecture and focus very much on a use case that is centered around customer support.

00:16:35.090 –> 00:16:41.949
Ranjit: So with that in mind, I wanted to talk about a few different high, level ways in which generative AI benefits Bpo’s

00:16:42.210 –> 00:16:59.730
Ranjit: so it benefits customers by giving them, you know. All the convenience. They need many different channels to come and interact. It gives them conversational capabilities so intelligent, you know, chat and things of that nature, and you’ll see a lot of that in the Demos that I’ll show you

00:16:59.880 –> 00:17:16.790
Ranjit: on the agent’s side when some, when an agent is assisting. There’s some administrative work that they have to do in terms of summarizing the call, and all of that. And then there’s also the actual work of helping the customer, and we provide tools that can assist the agent during the call

00:17:17.119 –> 00:17:24.970
Ranjit: by leveraging all the knowledge that’s been created and giving them suggestions on what they should tell the customer. And then finally.

00:17:25.180 –> 00:17:31.099
Ranjit: you know, as more and more calls go through, go through your call center and go through the platform.

00:17:31.180 –> 00:17:35.249
Ranjit: You know, there’s there’s what we call continuous learning

00:17:35.280 –> 00:17:47.310
Ranjit: and the ability for you to improve the the quality of the knowledge that you have in your enterprise. You know as you as you move forward the goal being that over time

00:17:47.650 –> 00:17:53.680
Ranjit: as more and more customers come in, they are able to resolve their inquiries without needing an agent for assistance.

00:17:56.870 –> 00:18:21.900
Ranjit: So on this slide, I’ve kind of shown the different modules of of our platform. Like, I explained, we have a few modules on the front end. You you can have a conversational experience. You can build an app like experience with a tool called a solo. We have what what I call a lightweight contact center application. And then on the back end we have Bpm, a low-code builder and a whole collection of analytics tools

00:18:22.190 –> 00:18:34.240
Ranjit: for today’s discussion. I’m going to focus on the front end. So just to be clear again, the focus is on front end conversational platform used for customer support.

00:18:36.510 –> 00:18:45.399
Ranjit: So now I’m going to kind of drill down into the specific topic of the webinar. So if you look at

00:18:45.700 –> 00:19:14.730
Ranjit: you know, bpo, it’s it’s been a somewhat manually oriented endeavor with with, you know, with agents helping customers, and and not necessarily having a completely mechanized and learning capability to, to, you know, to provide a better experience for customers over time. Now, with the advent of Gpt and and technologies like that. We see that there’s a big opportunity to have a paradigm shift when it comes to Bpo capabilities.

00:19:14.760 –> 00:19:19.679
Ranjit: And now everyone has experimented with

00:19:19.770 –> 00:19:38.840
Ranjit: you know, using Openai and using tools like that and made Api calls and been amazed at the quality of response you get from them. But the reality is that in order to actually deploy GPT. In A in a real enterprise setting, it needs to be bundled with a lot of infrastructure. That, you know, enables the

00:19:38.860 –> 00:19:56.699
Ranjit: the the tool to be enterprise friendly with with all the guardrails for security and learning and and integration with different tools and all of that. So we put together a bundle for specifically for Ppo. And here’s kind of a high, level summary of the features.

00:19:56.730 –> 00:20:12.409
Ranjit: So the first one, obviously is you leverage the capability of GPT. Both for search and for response. So when anyone comes in with a question. you can actually search your knowledge base, get the data you need. And then you can formulate a nice, user friendly response that you can send back to them.

00:20:13.490 –> 00:20:24.620
Ranjit: The second point is that these conversations can be made personalized. They can made be made contextual. People can ask follow-up questions. The third one is.

00:20:24.980 –> 00:20:35.390
Ranjit: you know, if you detect that at some point the conversational area is not meeting the needs of the customer, you have the ability to push that push the conversation to an agent who can assist

00:20:36.590 –> 00:20:45.770
Ranjit: Gpt is a generalized technology. So tools like Chat Gpt operate on the on general knowledge that’s available.

00:20:45.800 –> 00:21:04.940
Ranjit: You know, sort of globally. And so they can answer general questions. So we have have the ability to create customized versions of GPT. That are customized to your content. This could be, you know, public facing chat bots that are customized based on your website or internal facing chat bots that are customized, based on

00:21:05.240 –> 00:21:07.060
Ranjit: internal documents.

00:21:07.600 –> 00:21:15.490
Ranjit: As you use these Gpts and these Llm models more and more, you’ll find that cost becomes an interesting factor.

00:21:15.580 –> 00:21:23.599
Ranjit: The the standard pricing paradigm for these Llms is based on tokens.

00:21:23.800 –> 00:21:27.920
Ranjit: and if you start asking a lot of questions that have long answers.

00:21:27.960 –> 00:21:35.439
Ranjit: You’ll find that your tokens consumption gets high very quickly, and you’re you’re paying a large amount to to use these Llms.

00:21:36.220 –> 00:21:48.689
Ranjit: And then you every time you ask a question you want the ability for this. You know, for the response that comes back from your model

00:21:48.910 –> 00:21:57.210
Ranjit: to to be subject to curation if you need. So you need the ability to integrate with the curation workflow which I’ve kind of shown here. But you also need the ability

00:21:57.220 –> 00:22:06.320
Ranjit: to. You know, if if there is a really good response to keep it so that you can, you can reuse that response in the future. And then there’s a couple of other features around

00:22:06.560 –> 00:22:24.969
Ranjit: using, you know. Rule based approaches to to dealing with specific types of queries and putting guardrails in place to ensure to ensure accuracy, to ensure scope and things like that. So this is kind of a full featured bundle that is available right now with the autonomate platform.

00:22:25.620 –> 00:22:30.329
Ranjit: So with that enough of talking, and I’m gonna jump into Demos.

00:22:30.400 –> 00:22:35.410
Ranjit: So the first demo I’m going to show you is what we call the travel assistant.

00:22:35.690 –> 00:23:01.970
Renga: So let me just go. Just talk a little bit about this before I get into it, and just wanted to quickly point out to the group that feel free to drop in your questions. The QA. Module we would keep ample time at the end of session to answer these. So please keep posting your questions there, and we’ll get to it by the end of the session. Thank you. Or do you?

00:23:02.090 –> 00:23:17.030
Ranjit: Yeah, thanks, Ranga. Good point. I think you know, under normal conditions, we II wouldn’t mind if people, you know, had questions in the in in the middle of my presentation. But since we have a larger audience this time, it probably makes sense to move everything to a. QA. At the end.

00:23:17.730 –> 00:23:29.959
Ranjit: So, you know, make note of your questions. So the first Demo, I’m going to show you is what we call a travel assistant in this. What we’ve done is set up a custom version of

00:23:30.160 –> 00:23:40.629
Ranjit: GPT. That is built on a specific website. And in this case we picked the lonely planet travel website. All of the data is unstructured and it is public.

00:23:41.130 –> 00:23:47.970
Ranjit: The objective of this chat experience is to give informational responses to customers who have questions about travel.

00:23:48.050 –> 00:23:52.230
Ranjit: In this case, we we are using Gpt 3.5.

00:23:52.520 –> 00:24:10.260
Ranjit: And one of the main reasons is that the answers are long. They have lots of tokens, and so we make a trade off of whether I should pay a lot more to use. GPT. 4, which has incrementally better quality in some cases, or save a lot of money by using Gp 3.5.

00:24:10.600 –> 00:24:21.960
Ranjit: And under the hood, the technology that we use is is something that some of you may be familiar with. It’s called retrieval augmented generation. What this means is that.

00:24:22.470 –> 00:24:39.509
Ranjit: when we send the question to Openai, GPT. 3.5. We’re actually accompanying it with some contextual information that we get from a search process. And I’ll kind of explain that as we go forward. So let’s get get into the demo itself right now.

00:24:39.960 –> 00:24:43.839
Ranjit: So I’m gonna put put these other windows down for a second.

00:24:44.010 –> 00:24:47.310
Ranjit: I’ll keep this window open.

00:24:47.630 –> 00:24:49.880
Ranjit: cause we may. We’ll probably come back to it.

00:24:50.760 –> 00:24:54.240
Ranjit: and I’m going to talk to this travel assistant.

00:24:55.220 –> 00:24:58.519
Ranjit: So the travel assistant, like, I said, has been

00:24:58.570 –> 00:25:02.449
Ranjit: built with public data from the lonelyplanet.com website.

00:25:02.770 –> 00:25:10.810
Ranjit: and every time someone asks a query. So I’m going to type in my first query here, just going to copy and paste from another document.

00:25:14.800 –> 00:25:17.500
Ranjit: So when I ask this first query.

00:25:17.680 –> 00:25:21.729
Ranjit: what’s happening is that we run a search

00:25:21.840 –> 00:25:33.119
Ranjit: for information to get some fragments of information from the lonely planet.com website initially. And then we use that to create a prompt that we then send to GPT. 3.5.

00:25:33.150 –> 00:25:37.680
Ranjit: The benefit here is that you? You may have heard the term grounding.

00:25:37.700 –> 00:25:56.590
Ranjit: and we use grounding to ensure that the Gpt does not come back with misinformation, or, as some of you may have heard, hallucinations, so so so you ground the you know the engine, the model by feeding it, some information that’s authoritative

00:25:57.040 –> 00:26:10.149
Ranjit: and then based on that, it uses, of course, its capability to generate a nice looking response in terms of you know all, all of the reasons about you know the best time to visit Italy, and all of that. In addition.

00:26:10.430 –> 00:26:13.309
Ranjit: we also come back and give you a link.

00:26:13.760 –> 00:26:14.580
Ranjit: 2.

00:26:14.980 –> 00:26:25.220
Ranjit: The specific page, the top ranking page on the site from which a lot of this information was provided. So we tie it on both ends to make sure that the information is good and accurate.

00:26:26.040 –> 00:26:30.150
Ranjit: I have a few more queries that I’m going to punch into this chat board.

00:26:31.340 –> 00:26:40.110
Ranjit: In this case, I said, do I need a visa? Now? Technically, this is a follow up question. and I’m not in most chat bots.

00:26:40.230 –> 00:26:55.959
Ranjit: This question will not be recognized, because, you know, the system won’t know. You know, visa, what you’re talking about. But in this case we happen to know that since the previous question was about Italy? You know, we can preserve the context through different through through multiple questions.

00:26:56.240 –> 00:27:00.689
Ranjit: you can do it with with unstructured content. You can do this to some extent.

00:27:00.700 –> 00:27:05.320
Ranjit: but you can do it much better with structured content, as I’ll show you in a subsequent demo.

00:27:06.430 –> 00:27:13.989
Ranjit: And then I say, ant to visit Morocco, which means again, since contextually in my mind, I am talking about visas.

00:27:14.270 –> 00:27:19.990
Ranjit: The real question is, Do I need a visa to visit Morocco. That’s the question that’s being asked. And here.

00:27:20.070 –> 00:27:23.020
Ranjit: you know, I’m able to get to get a good response to it.

00:27:23.180 –> 00:27:32.120
Ranjit: So we maintain the context. And we’re able to give a sensible response to this. I’m going to start with a kind of a new line of questioning.

00:27:32.690 –> 00:27:34.060
Ranjit: and this is.

00:27:34.470 –> 00:27:51.500
Ranjit: you know, but what are the best travel destinations for for older people? This is almost, you know, the tool almost has to write an essay for me on this, and, as you can see, it does, it does a pretty good job. There’s a general query, and in response to the general query. It gave me, useful information that it

00:27:51.530 –> 00:27:54.670
Ranjit: collected from the pages of lonelyplanner.com

00:27:54.700 –> 00:28:02.729
Ranjit: up, and in addition, as before, it gave me you know, a link that it thought would be most effective

00:28:02.760 –> 00:28:10.429
Ranjit: or most relevant to the content that that will display. So I’m going to now get into a couple of specific questions

00:28:15.600 –> 00:28:18.359
Ranjit: about historic lighthouses in California.

00:28:18.480 –> 00:28:46.020
Ranjit: And again, I’m not gonna say a whole lot about this. It’s just that you know the the tool does a good job of understanding what I want, and pulling information from different pages to construct a response. All of these information, all of these responses I have requested. Be kind of more lengthy and informative. A. And in the next, in the next demo you’ll notice that I’ve I’ve requested very precise responses.

00:28:46.840 –> 00:28:53.180
Ranjit: This is another example of a very specific question, but it’s sort of a multi-part question. and

00:28:53.230 –> 00:29:06.540
Ranjit: you know you’ll you’ll see that I asked about. You know lullabout, Botanical Gardens, and a bunch of questions about it, and you’ll notice that the system responded with with, You know, to each of my questions properly.

00:29:07.040 –> 00:29:09.379
Ranjit: Okay. So now I’m going to pick.

00:29:10.420 –> 00:29:21.990
Ranjit: I’m typing in a question that says your website gave me incorrect advice, and I am stuck in Tunisia. This is an actual complaint from someone on the lonelyplanet.com website. And to showcase this a little bit, better

00:29:22.220 –> 00:29:29.230
Ranjit: keep. I’m gonna log in on the left side of my screen. I’ve this is our a 8 live product.

00:29:29.750 –> 00:29:34.000
Ranjit: and I’m going to log in and make myself available as an agent.

00:29:34.980 –> 00:29:37.490
Ranjit: So in this.

00:29:37.610 –> 00:29:47.199
Ranjit: you know, when when this query is typed in. what happens is that the system under the hood detects that there is a negative sentiment associated with this.

00:29:48.010 –> 00:30:01.169
Ranjit: and then what? What it does is it gives me kind of a generic response, but automatically transfers me to a live agent. So now, rather than me, talking to the virtual agent, you’ll see that I’m talking to the live agent.

00:30:01.330 –> 00:30:09.219
Ranjit: So in this manner, what happens is that based on in in this case, sentiment. It could also be based on the if. If the chat.

00:30:09.660 –> 00:30:19.600
Ranjit: what doesn’t understand, you know, it will transfer me automatically to a live agent. But in this case. you know, we use sentiment as the trigger.

00:30:20.090 –> 00:30:29.319
Ranjit: And so I just wanted to show you how that works. I’m not gonna say anything meaningful. But I’m gonna transfer the chat back alright. So the next thing I’m going to do here now

00:30:30.860 –> 00:30:40.030
Ranjit: is show you a different capability. So let me just reset this reset this conversation. I’m going to go back to the travel assistant

00:30:44.090 –> 00:31:11.659
Ranjit: and I’m I’m gonna show a capability where you can bypass Gpt and create your own rules. In this case you can think of this as being kind of an a upsell rule. So the question that the user asks is, what is the easiest way for me to apply for a visa. Yeah. Sorry to interrupt you, Ranjit the audience. One of the members of the audience is asking if you could probably zoom in a bit, especially into the chat window. I believe

00:31:11.720 –> 00:31:15.660
Renga: so, because it’s not so legible. Yeah.

00:31:16.630 –> 00:31:22.899
Renga: is that is that better? I think it looks better. We will wait to hear from that. Yeah.

00:31:23.060 –> 00:31:28.179
Ranjit: yeah. So I’ve I’ve just increased. I’ve just increased the size of the font.

00:31:28.730 –> 00:31:31.310
Ranjit: Hopefully, that’s helpful. Okay?

00:31:31.990 –> 00:31:39.589
Ranjit: So so now, in this case, when the user asks for for how to apply for a visa. Let’s say as a company. Let’s say I’m lonely. Planet.

00:31:39.680 –> 00:31:53.370
Ranjit: and I have a deal with Vfs. And Vfs, as you know, are are the, you know, well known. Visa processors. So what we do is we set up a rule that says, anytime someone asked about a visa. I want to show them

00:31:53.370 –> 00:32:09.740
Ranjit: the visa at your doorstep. Capability that Vfs go global provides. So in this case we we created a rule to bypass Gpt, and then just show you you can think of this as an ad or a sponsored link, or something of that nature. But that’s that’s what we’re showing here.

00:32:10.600 –> 00:32:20.219
Ranjit: Hopefully, this this is a you know, the font size is a little bit better now. And finally, I’m going to close with a question that’s completely out of scope

00:32:21.290 –> 00:32:25.980
Ranjit: and and again, the reason he is to show you here that we can set guardrails on

00:32:26.070 –> 00:32:37.260
Ranjit: you know the scope of the chat board so that anytime there’s a question that’s outside of scope. It will respond and and say that, look, you know, please, restrict your your questions to a particular topic.

00:32:37.790 –> 00:32:41.070
Ranjit: Okay. so this was demo number one.

00:32:41.390 –> 00:32:43.830
Ranjit: I’m gonna go back to my Powerpoint

00:32:45.470 –> 00:32:52.170
Ranjit: and show you demo number 2, which is using structured data.

00:32:52.420 –> 00:33:06.459
Ranjit: And in this case I’m gonna be talking to. You know, I have information in a database in a postgres database. This is private data. This is retail information about customers and products and orders and things like that.

00:33:06.560 –> 00:33:12.139
Ranjit: And the objective here is to have precise responses to specific customer queries

00:33:12.490 –> 00:33:32.330
Ranjit: in this, I’m using the Openai, GPT. 4 as my back end. Since there’s a few, a number of tokens, I’m not spending a huge amount of money. And then the approach we’re using here is is a combination of See, you know, generating SQL on the fly and what we call it’s it’s broadly prompt engineering. But the idea is to create dynamic prompts

00:33:32.990 –> 00:33:43.140
Ranjit: based on the questions that the that the users are asking. And and again, I I’m just giving. If for for those of you who are technically minded, you know this may you may find this interesting.

00:33:43.460 –> 00:33:47.099
Ranjit: so let me reset the conversation here.

00:33:48.060 –> 00:33:49.020
Ranjit: and

00:33:50.430 –> 00:33:52.950
Ranjit: I’m going to connect to the retail assistant.

00:33:54.170 –> 00:33:57.739
Ranjit: Let me see if I can increase the size of this winter.

00:33:59.890 –> 00:34:02.509
Ranjit: Okay, that’s that’s better. Actually.

00:34:02.700 –> 00:34:05.780
Ranjit: okay, so hopefully, this window is a good size. Now.

00:34:06.040 –> 00:34:09.400
Ranjit: Okay. So for the retail assistant, I

00:34:09.489 –> 00:34:22.459
Ranjit: you know the the way we’d set up the demo is that the Demo asks me to give give the name of a random customer, but in in this case I’m I’m going to ask of, you know. Get the name of a specific customer who, I know, has.

00:34:22.880 –> 00:34:25.449
Ranjit: you know, certain buying

00:34:25.820 –> 00:34:36.889
Ranjit: certain data so that I can. I can have a predictable flow for the demo. So I’m introducing myself as Joan Kowsack, and it now, what I’m going to do is show you another screen.

00:34:37.480 –> 00:34:40.049
Ranjit: That shows you the backend of this database.

00:34:41.980 –> 00:34:51.780
Ranjit: To describe what’s what’s in the database. So? Here we have a a database which has customers, and in this database we have, with 37,000 customers.

00:34:51.820 –> 00:35:05.890
Ranjit: About half a million orders, and about a million items that have been ordered. There are about 2,500 products and 7,700 stores. So this is the information in the database, and let me go into full screen mode here.

00:35:05.940 –> 00:35:09.480
Ranjit: and I’m going to just show you

00:35:09.540 –> 00:35:16.150
Ranjit: the the real information in the database about this this particular person.

00:35:16.620 –> 00:35:17.990
Ranjit: so that

00:35:19.560 –> 00:35:39.050
Ranjit: so that you can see what you know. So that you have context when I type in these questions. So these are all the orders that Joan Cozac has made. The most recent order, had an order date of 20 third of October. This shows the ship date. So anything with a ship date, I mean estimated ship date has not been shipped.

00:35:39.650 –> 00:35:44.720
Ranjit: Anything with an estimated delivery date has not been delivered. This shows that you know

00:35:45.580 –> 00:35:51.960
Ranjit: the you know the the you know the 4 items in this order. This was the quantity.

00:35:52.580 –> 00:36:08.180
Ranjit: These were the items that were ordered, and you know it. It describes each of these items in terms of color and size and price, and all of that. So just keep this in mind as I ask as I’m I’m gonna be Joan Kozak, and I’m going to ask a few questions.

00:36:08.510 –> 00:36:09.290
Ranjit: So

00:36:10.920 –> 00:36:13.349
Ranjit: let me type in first question.

00:36:13.940 –> 00:36:22.519
Ranjit: when will my most recent order be shipped. Unlike the previous example in this case, you really don’t need long answers. You need very precise answers.

00:36:23.290 –> 00:36:29.399
Ranjit: You’ll see that there’s an estimated ship date of 1027, and the tool was able to actually pull this off.

00:36:30.900 –> 00:36:32.520
Ranjit: Let me ask the next question

00:36:33.670 –> 00:36:38.919
Ranjit: again. I am sorry I’m with the delay I’m copying and pasting from a different page here.

00:36:39.130 –> 00:36:39.920
Ranjit: oops.

00:36:42.800 –> 00:36:47.680
Ranjit: when will it be delivered? There’s an estimated ship, date of

00:36:47.720 –> 00:36:52.509
Ranjit: 1102 again. This is sorry for the American style dates.

00:36:52.720 –> 00:37:00.449
Ranjit: So you you’ll see that it’s it’s able to maintain the context that I’m talking about the most recent order, and it’s able to tell me when that’s going to be shipped.

00:37:01.610 –> 00:37:02.960
Ranjit: So now

00:37:03.140 –> 00:37:12.029
Ranjit: let’s say I come in and say something like, Oh, no, I have to attend a wedding on that day. Now to a normal chat board, this this makes no sense.

00:37:12.220 –> 00:37:20.800
Ranjit: but if you think about it in the terms of of a conversation. What I’m really implying is that on that day I’m I won’t be around for the order to be delivered

00:37:20.900 –> 00:37:23.700
Ranjit: so so to to

00:37:24.030 –> 00:37:27.469
Ranjit: to handle situations like this. The Chatbot needs to

00:37:27.740 –> 00:37:43.860
Ranjit: have some sort of inferencing or reasoning capability, and you, as you can see, the Chat port, was able to do that, and based on what I said, it was able to figure out that I wanted to reschedule my my shipment. I’m not gonna get into the rescheduling. But I just wanted to show that

00:37:44.090 –> 00:37:47.379
Ranjit: we have this inferencing capability as well.

00:37:48.460 –> 00:37:49.290
Ranjit: Okay.

00:37:49.810 –> 00:37:53.949
Ranjit: so I’m asking what is in the order and let me come back here to the screen.

00:37:54.020 –> 00:37:56.110
Ranjit: You’ll see that in the order there is.

00:37:56.200 –> 00:38:00.170
Ranjit: you know, there are 4 items, a fabricam contoso

00:38:00.680 –> 00:38:05.440
Ranjit: theater, and you know, a bunch of different things like that. And as you can see.

00:38:06.400 –> 00:38:15.170
Ranjit: the system was able to pull this. And this is basically the ability to do data itemization, you know, by running a SQL. Query in the backend.

00:38:19.040 –> 00:38:22.550
Ranjit: how much did I pay? And up

00:38:22.760 –> 00:38:27.980
Ranjit: now this information is not available in the database? It requires the engine to actually do

00:38:28.000 –> 00:38:35.340
Ranjit: to compute. based on the quantity and the price the engine actually needs to compute and

00:38:35.380 –> 00:38:38.880
Ranjit: add up the the total in, you know.

00:38:39.110 –> 00:38:42.070
Ranjit: pricing and and tell me the total that I paid.

00:38:43.400 –> 00:38:44.260
Ranjit: Okay?

00:38:45.340 –> 00:38:47.550
Ranjit: Again, going back

00:38:48.250 –> 00:38:57.200
Ranjit: again to so show that I’m maintaining context. So I asked this question like 7 questions ago, and I’m still able to maintain context.

00:38:58.930 –> 00:39:02.449
Ranjit: let me ask another question.

00:39:07.450 –> 00:39:08.280
Ranjit: So

00:39:08.430 –> 00:39:11.039
Ranjit: you know. So so here

00:39:11.280 –> 00:39:19.890
Ranjit: we you know we we don’t. I. I’m just asking questions that are out of scope for the chat bot, and it’s it’s not able to provide any answers on it.

00:39:20.600 –> 00:39:22.210
Ranjit: So I just wanted to show you that.

00:39:22.540 –> 00:39:27.890
Ranjit: How many orders have I placed so far?

00:39:29.100 –> 00:39:31.350
Ranjit: And you know it’s

00:39:31.530 –> 00:39:40.520
Ranjit: it. It gives me the proper number and things like that. So so this hopefully gives you an idea of our capability

00:39:40.850 –> 00:39:46.720
Ranjit: to ask. You know, very specific questions from a retail database creating something like this.

00:39:47.080 –> 00:39:50.600
Ranjit: I should also say that to create the first web. So you know, Web.

00:39:50.630 –> 00:39:54.700
Ranjit: Base Chatbot! Took us 2 or 3 days

00:39:55.050 –> 00:39:58.790
Ranjit: creating something like this. If we have access to your database on your schema.

00:39:58.820 –> 00:40:04.710
Ranjit: we can build a Chatbot that an intelligent chat Bot based on your data

00:40:04.930 –> 00:40:10.539
Ranjit: within, you know, 3 to 5 days. So I’m going to see if there are any other interesting questions

00:40:11.860 –> 00:40:14.270
Ranjit: that I can ask. Let’s see.

00:40:16.520 –> 00:40:19.340
Ranjit: Okay. let me

00:40:21.160 –> 00:40:23.540
Ranjit: plus this. And this is basically to do.

00:40:23.900 –> 00:40:31.679
Ranjit: You know, if you’re familiar with SQL to kind of get get the maximum value from from a particular column.

00:40:32.110 –> 00:40:34.019
Ranjit: and if you look at the pricing.

00:40:35.610 –> 00:40:49.009
Ranjit: so it looks like this is the pricing column. 600 is probably the highest price that we have. So it’s done a good job. In addition, it told me when I bought these items as well.

00:40:49.780 –> 00:41:02.349
Ranjit: And then I’ll I’ll close with kind of a question that’s a little bit hard. because even writing a SQL. Query here, for this is hard, because you have to do some sort of aggregation and grouping.

00:41:03.350 –> 00:41:14.110
Ranjit: So what categories of products have I purchased? You’ll see that for all the products there’s a category column here. So the system is able to tell me

00:41:14.120 –> 00:41:23.110
Ranjit: and and do kind of a grouping of all of this data and give me a summary of you know what I purchased. I’m gonna try a couple of things here.

00:41:23.710 –> 00:41:28.300
Ranjit: which is to see if I can break the privacy guidelines.

00:41:32.860 –> 00:41:44.019
Ranjit: I’m trying to ask and see if I can figure out what other people have bought. A. But the system has these privacy guidelines built into it, so it’s not able to violate any

00:41:44.220 –> 00:41:51.509
Ranjit: it. It tells me that I cannot violate the privacy guidelines and give you information that’s outside of what you have purchased.

00:41:52.900 –> 00:42:01.629
Ranjit: And then I am going to ask a totally out of scope question, and the system should respond that you know I am not able to do this because this is out of scope for me.

00:42:01.680 –> 00:42:04.389
Ranjit: So so this sort of concludes,

00:42:14.590 –> 00:42:18.770
Ranjit: okay, lemme lemme check here. Okay, so we’re we’re doing okay on time.

00:42:21.260 –> 00:42:32.050
Ranjit: Okay? Great. So I showed you this one I showed you this demo.

00:42:32.350 –> 00:42:36.589
Ranjit: The third one I want to show you is something that’s a a little bit different.

00:42:36.890 –> 00:42:45.269
Renga: And before you get started with the third demo I just wanted to check. If the audience is doing better when it comes to being able to make out

00:42:45.460 –> 00:42:57.250
Renga: the sequence of events on screen. Would you like it to be a little more larger, because there are a couple of windows at the same time? If you could just give me a thumbs up or answer in the QA. That would be great.

00:42:59.010 –> 00:43:05.660
Ranjit: Okay? So yeah. So this this demo that I’m going to show you hopefully, it’s it’s not chat based.

00:43:05.720 –> 00:43:09.660
Ranjit: So it it should. You know the the visibility should be better

00:43:10.350 –> 00:43:16.359
Renga: perfect. You can go ahead, and this seems to be no response of it, as you’ve known use is good news.

00:43:16.620 –> 00:43:21.459
Ranjit: Okay? Great so this demo I’m gonna show you is what we call field extraction.

00:43:21.490 –> 00:43:32.250
Ranjit: And let me explain what we do here. So what what we have is we’re using email as a channel. and we have the ability for a user.

00:43:32.840 –> 00:43:50.139
Ranjit: Let’s say, this is your customer who wants to send you an invoice or a bank statement, or something like that? This customer can send you a Pdf document on email. And what we can do is we can extract specific fields from this email and fill them in automatically, fill them into a form on the back end

00:43:50.880 –> 00:44:06.600
Ranjit: for this we’re using. You know what we call we, we’re using the Google Palm, 2 model and specific the palm 2, you can say Llm, and within that, specifically, the text Bison model. And the approach here is somewhat little bit more traditional.

00:44:06.890 –> 00:44:13.779
Ranjit: We’ll be using just a vector store with embeddings and things like that to to extract the fields from this Pdf document.

00:44:14.720 –> 00:44:17.989
Ranjit: So let me show you this, demo.

00:44:18.010 –> 00:44:20.210
Ranjit: And for this, what I’m going to do

00:44:20.790 –> 00:44:33.680
Ranjit: I don’t need all these windows. I’m just gonna go to this window where I’ve logged in as a you know, as as a particular user. I’m going to compose an email to

00:44:34.650 –> 00:44:37.710
Ranjit: our email gateway. And I’m going to say

00:44:40.050 –> 00:44:45.530
Ranjit: something like, Here is my invoice. And I’m going to attach a file.

00:44:46.350 –> 00:44:51.340
Ranjit: sample, invoice.

00:44:52.460 –> 00:44:53.450
Ranjit: open

00:44:57.000 –> 00:45:04.230
Ranjit: and send it. So here I am, just a regular user. And I’m sending a sample invoice by email. So I’m going to send this.

00:45:04.260 –> 00:45:09.620
Ranjit: And now on the back end. I wanted to show you what we do with with these incoming invoices.

00:45:10.910 –> 00:45:13.719
Ranjit: So we have a backend process

00:45:13.980 –> 00:45:18.710
Ranjit: that looks at incoming invoices, and then it extracts the fields from them.

00:45:18.760 –> 00:45:30.190
Ranjit: So it analyzes the invoice and extracts the fields and puts them into a into a into a form. And and II just wanted to show you what that form looks like. So let me just

00:45:30.580 –> 00:45:35.270
Ranjit: log in on the back end. I should see the form coming in.

00:45:36.450 –> 00:45:39.679
Ranjit: Let me just log out and log back in just to

00:45:42.370 –> 00:45:49.890
Ranjit: just to make sure I have the right data. and I’m gonna click on all tasks. and you’ll see that a few seconds ago

00:45:50.000 –> 00:45:56.840
Ranjit: a new task came up to me at the back end. So this task has been triggered by some one emailing an invoice.

00:45:56.870 –> 00:46:01.060
Ranjit: So when I click on this task, you’ll see

00:46:01.860 –> 00:46:13.459
Ranjit: that this contains a bunch of details on the invoice, and I’m going to just show you what the invoice looks like, so that you have an idea. so this is the sample. Invoice.

00:46:15.250 –> 00:46:19.960
Ranjit: So it’s able to. you know. Pick up the the customer name.

00:46:20.280 –> 00:46:23.600
Ranjit: Yeah. The the item which is brochure design

00:46:23.620 –> 00:46:30.189
Ranjit: the Val. The amount of the IGST. The amount of the invoice assess, and all of these kinds of things. So

00:46:30.540 –> 00:46:37.110
Ranjit: it’s an automated way for customers to mail you structured documents, and then you can actually look at them.

00:46:37.390 –> 00:46:59.359
Ranjit: You know, you can actually extract the information and save a lot of time by having them stored in the back end. You can do many different things like this. You can automate auto populate forms. You can auto populate crms, you can auto populate ticketing systems like service now, and you know things of that nature. So you can do a whole bunch of interesting things with this field extract. I’ll show you another example here

00:46:59.500 –> 00:47:05.240
Ranjit: where, as the user, I’m gonna same user is, gonna send another document.

00:47:07.290 –> 00:47:11.639
Ranjit: And in this case, here is my

00:47:11.660 –> 00:47:13.280
Ranjit: bank statement.

00:47:17.930 –> 00:47:20.349
Ranjit: and I’m going to attach the bank statement

00:47:22.220 –> 00:47:36.510
Renga: and Regent, just to let the audience know that there are Christians already coming in in the QA. Section. But we’re so I have another, maybe 5 min of demo. And then I can get to the questions

00:47:36.630 –> 00:47:37.580
Renga: perfect.

00:47:38.080 –> 00:47:39.210
Ranjit: Okay.

00:47:39.550 –> 00:47:43.879
Ranjit: so I’m gonna attack. Oh, oh, I already did that. Sorry. Okay, so this.

00:47:47.890 –> 00:47:52.869
Ranjit: So I’m going to send the bank statement just to show you how things work on the back end.

00:47:52.920 –> 00:47:56.900
Ranjit: We have an email handler that is looking for incoming emails.

00:47:57.220 –> 00:48:00.769
Ranjit: I’m waiting for a new email to show up now.

00:48:01.100 –> 00:48:11.340
Ranjit: You know a A again rather than just have you guys wait. I thought, it’s useful to show you what’s going on behind the scenes. So we have an email gateway that receives the

00:48:11.450 –> 00:48:15.259
Ranjit: bank statement. And then when I click on the bank statement.

00:48:15.440 –> 00:48:27.440
Ranjit: it shows me that it’s actually sent it to the back end here. So again. I’m just, you know, showing you some of the under the hood pieces just in case you’re interested.

00:48:27.510 –> 00:48:34.200
Ranjit: Now let me come back here. So here I am waiting. And I got my previous

00:48:34.560 –> 00:48:38.169
Ranjit: a document about 3 min ago. I’m just gonna refresh this page.

00:48:43.720 –> 00:48:47.070
Ranjit: Okay. By the way, this is all

00:48:47.150 –> 00:49:00.779
Ranjit: training data and things like that. So I’m not exposing any any confidential information to you. So you’ll notice that a new fee, a a a new, a new task came in a few seconds ago. And in and in this case what we’ve done is

00:49:00.870 –> 00:49:06.060
Ranjit: we recognize that this is a bank statement. And I’m gonna show you what that bank statement looks like.

00:49:07.870 –> 00:49:13.560
Ranjit: And and, by the way, this is, you know, something we’ve taken and modified

00:49:13.700 –> 00:49:30.869
Ranjit: so that there’s no confidential information. But this bank statement has has a lot of details. And this so, for example, I’m asking for salary credit. and you’ll notice that there is nothing obvious about salary credits except something that’s buried deep down in the document somewhere.

00:49:31.190 –> 00:49:41.770
Ranjit: Hopefully, I’ll find it. That shows that there’s a salary credit of 12, you know, 12.6 K somewhere here. So I’ll I’ll look well. Ok, let me just do

00:49:44.990 –> 00:49:47.740
Ranjit: Okay.

00:49:51.850 –> 00:49:54.259
Ranjit: Pdf, search is slow.

00:49:54.600 –> 00:50:06.340
Ranjit: Okay, here. So you, you’ll notice that there’s a bulk posting by salary of 12.6 K. And we were able to actually identify that in this large table and extract it and put it into a field.

00:50:06.460 –> 00:50:09.759
Ranjit: Likewise we can, you know. Sorry about the

00:50:09.990 –> 00:50:21.209
Ranjit: you know the column width here, but what what we have is the ability to extract information from pretty complex documents, and to auto populate these fields with them.

00:50:21.870 –> 00:50:30.439
Ranjit: So that’s one thing. And then the last item I wanted to show you was the capability for us to actually read documents

00:50:30.470 –> 00:50:37.020
Ranjit: read Pdf documents and run searches on them. So what I’ve done is, I’ve taken a collection of

00:50:37.360 –> 00:50:48.999
Ranjit: Hyundai manuals and created a simple chat bot against which which you can ask questions. I’m going to show you what the manual looks like. And again, this is just an illustration

00:50:49.050 –> 00:50:51.060
Ranjit: of our ability to

00:50:51.650 –> 00:50:53.370
Ranjit: take

00:50:53.520 –> 00:51:01.129
Ranjit: Pdf documents. I showed you structured data. I showed you website content. But I also wanted to show you a simple capability with with Pdf.

00:51:01.370 –> 00:51:04.290
Ranjit: so again, I’m just gonna

00:51:04.490 –> 00:51:07.430
Ranjit: copy and paste couple of questions here.

00:51:13.250 –> 00:51:15.620
Ranjit: And this is something that’s on page

00:51:16.730 –> 00:51:18.270
Ranjit: 3, 63.

00:51:20.790 –> 00:51:21.670
Ranjit: So

00:51:22.380 –> 00:51:31.209
Ranjit: evasive steering assist function is described here, and you’ll see that we talk about the evasive steering assist function. I’m also going to ask a follow up question.

00:51:34.420 –> 00:51:36.300
Ranjit: Does it work for pedestrians?

00:51:38.460 –> 00:51:39.600
Ranjit: And it says

00:51:39.890 –> 00:51:47.000
Ranjit: it. It works for pedestrians. It helps avoid collisions with vehicles, pedestrian cyclists in the same lane. So you can see that it’s

00:51:47.430 –> 00:51:55.740
Ranjit: disinformation is from from this Pdf document I can. I can let me try one more

00:51:56.420 –> 00:51:59.199
Ranjit: question here. When ask a new question.

00:52:02.580 –> 00:52:07.439
Ranjit: I’m going to go to page 1, 17, where this is from.

00:52:10.390 –> 00:52:19.020
Ranjit: and we can talk about adjusting the front seat headrest and things like that. And you’ll see that this information is available from this page.

00:52:19.190 –> 00:52:22.590
Ranjit: I’m going to ask a couple of follow-up questions here.

00:52:25.920 –> 00:52:28.999
Ranjit: Can I recline the seat back toward the front?

00:52:29.800 –> 00:52:41.009
Ranjit: And it says, yes, you can do this, and how it tells me how you can do this, based on the information that’s in this document somewhere on this page. But I can ask us even more.

00:52:41.500 –> 00:52:43.249
Ranjit: you know, you can say

00:52:43.390 –> 00:52:56.630
Ranjit: specific question, what happens if I do this when my headrest is raised. So I want to specifically point out this point situation, where, if I raise the headrest and move my seat in front. Then it can actually hit the

00:52:56.810 –> 00:53:02.109
Ranjit: the windshield, or or the visor, or something like that.

00:53:03.100 –> 00:53:22.600
Ranjit: So it says, if you recline the feedback with the head restraint that may come in contact with the sun wears, or other parts of the vehicle again. This is just very illustrative. I just wanted to show you the ability to to look at many different types of documents. And and this kind of concludes what I wanted to show you from from a demo perspective.

00:53:22.870 –> 00:53:24.759
Ranjit: How are we doing on time, Ranga?

00:53:25.810 –> 00:53:40.859
Renga: We are doing pretty good on time, in fact. We wrapped up in 45 min. You want to take up the questions, or are you? Yeah, II just want to go. I I’m gonna spend 2 min on on a few just to make 2 or 3 key points. Here.

00:53:40.960 –> 00:53:43.300
Ranjit: I’m gonna come back to my deck here.

00:53:43.670 –> 00:54:02.420
Ranjit: and I wanted to talk about broadly. You know how you operate or how you operationalize Llms. It’s okay to have your teams, you know, experiment with Llms and things. But when you actually think of make, you know, operationalizing them and putting them in your enterprise. You have to you. There are many, many factors that you have to think of.

00:54:02.560 –> 00:54:04.950
Ranjit: So I wanted to

00:54:05.180 –> 00:54:18.240
Ranjit: layout a definition here. If you, if you can see this on the on the Powerpoint in in investing, there is a concept called efficient frontier. What this means is that you know you keep adjusting your portfolio so that you maximize.

00:54:18.540 –> 00:54:28.159
Ranjit: Your return for a fixed or a lower amount of risk. So you you can do that. And you want to kind of do the same thing with Llms.

00:54:28.320 –> 00:54:39.730
Ranjit: So we are building similar to. You know, to to what they do in in investing. We are building an efficient frontier idea for Llms that will continuously balance 4 variables

00:54:39.830 –> 00:54:57.470
Ranjit: security, cost, quality and speed. Now, you might say, you know, security you always want, say, security. But that’s not necessarily the case. If you’re just looking at public data. You really don’t care that much about security. But obviously, if you’re looking at human data, personal data, you care huge amount about security

00:54:57.670 –> 00:55:13.599
Ranjit: and so you know, again, I wanted to talk. We we have a bunch of different techniques like text redaction. If someone types in you know a credit card number or any other personal information that gets redacted from any

00:55:13.790 –> 00:55:27.479
Ranjit: string that is sent to the third party engine. We also use a technique. And I want to introduce this term to to all of you, because it’s it’s a very cool idea which is called homomorphic encryption.

00:55:27.670 –> 00:55:47.110
Ranjit: What this really means is that rather than send data, so rather than send my name, Ranjit Pradhmanabun to a third party engine who will know that that I exist. I’ll change my name to something else. I’ll change my age to something else, or and I’ll change my zip code to something else. So homomorphic means something that looks the same as

00:55:47.310 –> 00:56:13.569
Ranjit: as before, but has a different form. And so it’s like creating aliases. So you create a whole bunch of aliases that you sent to the third party engine, and then, when the data comes back, you remap the alias back to your actual identity. And so, W. What you’ve done is you’ve prevented information from leaking to the third party system again. I won’t go too deeply into this deck, but I wanted to talk about that specific point. And finally.

00:56:13.700 –> 00:56:15.650
Ranjit: you know, cost

00:56:15.670 –> 00:56:19.470
Ranjit: just keeps adding up. So we have many different techniques to keep costs low.

00:56:19.570 –> 00:56:30.170
Ranjit: So, for example, if you like a response that comes back from an Llm. We can cache it so you can start caching it and start. You don’t have to go back to the Llm. Someone asks if anytime someone asks the same question

00:56:30.190 –> 00:56:55.889
Ranjit: under normal customer support situations you have, you know what we call the long tail. So there’s probably 50 to a hundred questions that people are asking all the time. And if you just cash good responses to those questions, either from an Lm, or maybe even human curated. Then you don’t need to go incur the cost of calling a third party you know tool. So I’ll you know that’s all I wanted to say about that.

00:56:55.950 –> 00:56:59.999
Ranjit: And so I’ll yeah. I’m ready to take questions now.

00:57:01.580 –> 00:57:21.720
Renga: Awesome. Thank you, Ranjit. And actually the first person might seem like a good segue to what you just described at the last as what are the skills that required to develop such pods? I don’t know who submitted this, but I hope you could provide additional clarity if required. But, Runjit, could you answer that first question if you can.

00:57:21.770 –> 00:57:24.639
Ranjit: Yeah. Ii think so.

00:57:25.280 –> 00:57:32.119
Ranjit: The to be honest, the person most suit that when you create a bot, there’s basically 2 things. One is

00:57:32.180 –> 00:57:41.500
Ranjit: the flow the content. And for that it’s the you know, the the business user can create these things. Let me show you what the tool looks like.

00:57:42.460 –> 00:57:51.360
Ranjit: It’s a you know. It looks a little bit crazy when you look at it initially, but II want to guarantee that it’s a

00:57:52.120 –> 00:57:55.190
Ranjit: so when you create chat dialogues.

00:57:56.090 –> 00:58:02.139
Ranjit: This is what it looks like. You can drag and drop these different blocks to to tie together the different pieces of your dialogue

00:58:02.160 –> 00:58:10.189
Ranjit: and most of the dialogues are business friendly, so a human can. I mean, not a human. A business person can actually

00:58:10.250 –> 00:58:14.890
Ranjit: go and create these dialogues. But anytime you need an integration. So let’s say

00:58:15.030 –> 00:58:36.700
Ranjit: you want to integrate with salesforce. You might need it help. Now we we try and address that problem by giving you a whole bunch of connectors right upfront. So you can just use the existing connectors. Or if if you then need to create an integration with the tool that we don’t have a connector for. You will need it. Help. And someone would go to the back end on a tool like this

00:58:37.260 –> 00:58:44.590
Ranjit: the and they’ll have to write what we call services. And these services are just. For example.

00:58:45.390 –> 00:58:53.679
Ranjit: they look like this, they are templatized, but they need a little bit of knowledge of Javascript, so kind of a long answer to the question hopefully. It was useful.

00:58:56.740 –> 00:59:20.279
Renga: Sure, I hope, that answers the question. I wouldn’t know who has that. So who, if you’re listening, if you can let us know that, answer the question, or you have a follow up to that one. Please let us know. The next question is from Alex, and this mainly to the morocco. Use case that you talked about. He asked. If that is just about the visa, or was it about what could you do there when I

00:59:20.530 –> 00:59:26.530
Renga: see the question? I believe it’s about asking, what could you do in Morocco, if you’re allowed

00:59:26.560 –> 00:59:31.729
Renga: permissible, stay there Alex, let me know if I frame your Kristen correct

00:59:31.910 –> 00:59:34.340
Ranjit: when when I when I did the demo

00:59:34.490 –> 00:59:40.699
Renga: when I did the demo, it was specifically about visas. It was I was using that as a follow up question.

00:59:41.010 –> 00:59:48.100
Ranjit: But so let me just try and answer that really quickly. So when I started up this travel assistant

00:59:49.060 –> 01:00:01.310
Ranjit: you know, I started out with a few questions about Italy, and then III said, Do I need a visa? So so I actually, it’s it’s worth playing this back again, because this is interesting.

01:00:01.380 –> 01:00:14.039
Ranjit: So I’m you know, I ask a question about what is the best time of the year to visit Italy, which is fine. This is a fully formed question. so the system is able to understand that and and ask, ask a question about it

01:00:14.540 –> 01:00:15.250
Ranjit: up.

01:00:15.980 –> 01:00:33.490
Ranjit: And then the second question I asked is, Do I need a visa. Now, the implication here is that since I’ve already asked about Italy, that context is you know whether I need a visa to visit Italy. I mean, we need to understand the context to answer this question, and, as you can see, we do that. The third question I asked was.

01:00:33.530 –> 01:00:43.979
Ranjit: and to visit Morocco. So because I asked about a visa in my previous question, this next question is about whether I need a visa to visit Morocco. I mean, this is conversational style.

01:00:44.440 –> 01:00:55.840
Ranjit: And so in, you know, when I showed you the demo, that’s what I was showing the ability to preserve context. And in this case. It’s it’s it’ll give you a long answer about. You know all you need to do.

01:00:56.110 –> 01:01:04.599
Ranjit: Travelers from these countries don’t need visas, and then you know a whole bunch of things about that. So, but if you want to have a specific

01:01:06.640 –> 01:01:21.190
Renga: and Ranjith, if it’s possible, would you just mind increasing the size of this window. Or you could minimize the other window, too, so that this takes center stage

01:01:21.510 –> 01:01:22.270
Renga: him.

01:01:22.490 –> 01:01:23.290

01:01:30.210 –> 01:01:31.010
Ranjit: okay.

01:01:33.890 –> 01:01:36.359
Ranjit: so it talks about, you know, all, all the different.

01:01:38.730 –> 01:01:43.360
Ranjit: Oh, you know what II in this case, II should have switched context.

01:01:43.440 –> 01:01:44.610
Ranjit: So let me see.

01:01:52.850 –> 01:01:53.620
Ranjit: okay.

01:01:55.040 –> 01:01:56.980
Ranjit: is this visible? Hopefully.

01:01:58.190 –> 01:01:59.280
Renga: this is.

01:01:59.390 –> 01:02:00.120
Ranjit: yes.

01:02:01.000 –> 01:02:15.339
Ranjit: so I can ask, just a direct question on what activities I can do in Morocco, or you know, or I can ask ask it as a follow up question. So all of these are possible. Generally speaking, we do a good. We do a reasonable job

01:02:15.470 –> 01:02:18.670
Ranjit: where anytime you’re using these informational responses.

01:02:18.750 –> 01:02:37.900
Ranjit: The follow ups are a little bit trick, you know, tricky, we you know. So what we do is we focus on just giving the best you know the full full responses or like I showed in the other demo, we can specifically say, are you asking a follow up question, or are you asking a new question? And then the user can, you know, supply the information they need?

01:02:39.610 –> 01:02:43.040
Ranjit: Sorry, long answer, Alex. Sorry about that, but hopefully it was useful.

01:02:43.620 –> 01:03:03.670
Renga: No, Alex also had a follow up question, and I believe this is for the case where you talked about the orders for for Joanne, and his question is about the neighbor asked, could you provide a name that exists in the database? I think it is mainly to find out if you know your neighbor. And if you would probably want to ask about them.

01:03:04.460 –> 01:03:12.410
Ranjit: Yeah, I can ask. But you know, so let me let me try that. I don’t know if I know any neighbors. But

01:03:12.450 –> 01:03:15.870
Ranjit: let me ask you know, let me try something here.

01:03:16.500 –> 01:03:22.950
Ranjit: I haven’t tried that. So this is all live and dangerous thing for me to do in the demo. But I’m gonna try it. Anyway.

01:03:25.210 –> 01:03:27.900
Ranjit: I’m going to go in as Joan M. Kozak.

01:03:29.430 –> 01:03:31.639
Ranjit: And I am going to ask, what did

01:03:33.770 –> 01:03:35.660
Ranjit: Michelle purchase?

01:03:48.960 –> 01:03:59.389
Ranjit: And I’m gonna pray that it does the right? Okay? Good. Yeah. So so me, Michelle is a legitimate Ca, you know, these are all legitimate customers in the database.

01:03:59.400 –> 01:04:03.150
Ranjit: And so, as you can see it, you know it. It prevents,

01:04:03.250 –> 01:04:06.860
Ranjit: Jones inquiries from asking about what Michel purchased.

01:04:08.210 –> 01:04:26.650
Renga: Awesome. I hope that answers your question, Alex. Let us know if you have any follow ups on that one. I’m moving on to the next one from Gunjin. the question is, can the system also capture and extract unstructured data like something which is handwritten on a physical account opening form in a bank.

01:04:26.810 –> 01:04:30.770
Ranjit: Yeah. It. It can do that. There’s

01:04:31.440 –> 01:04:42.460
Ranjit: a limit to what it can do so we we. So in this case. So what I did was I was able to extract Pdf from documents. But this was actually, you know. Text. Pdf.

01:04:42.480 –> 01:05:02.800
Ranjit: now, let’s say there’s an image. We have the ability to do Ocr on images and extract text. But if it’s really badly handwritten, or something like that, the quality of the results will be will be poor. But we do have. An Ocr engine, and we have the ability. We we do. We have some other tools that can, for example.

01:05:02.980 –> 01:05:15.509
Ranjit: you know, do Ocr on images of identity, id cards like other card pancard and things like that, pull information and stick it into the into, you know, into your flows, and into your forms and and things of that nature.

01:05:15.860 –> 01:05:16.710
Ranjit: So

01:05:16.970 –> 01:05:28.079
Ranjit: yeah, so so so, not not a good, very specific answer, Benjamin. Sorry about that. But it, it really varies by quality. And one more thing I’ll point out is that

01:05:28.320 –> 01:05:34.940
Ranjit: you know, for example, signature, matching and tools of that nature. Those are kind of hard. Because,

01:05:35.200 –> 01:05:41.529
Ranjit: it’s really hard to get enough accuracy to make the tool useful. So I’ll just throw that out there also.

01:05:43.110 –> 01:05:59.289
Renga: Awesome. I hope that answers your question. Good Jen. We are moving on to the next one. I believe it was asked at the point in time during the presentation, and it’s anonymous, so I wouldn’t know whom to direct this query to. But they asked, Is there a way one could try this?

01:05:59.330 –> 01:06:15.610
Renga: So I don’t know what this question means specifically, to which point in the presentation, but I will let them post a follow up. There’s another question about, where are these Lms hosted? And how’s the data that we share with these Lms secure.

01:06:16.160 –> 01:06:19.490
Ranjit: Okay, so in.

01:06:19.700 –> 01:06:26.780
Ranjit: You know, in the Demos that I showed you the the let me go back and show my Powerpoint.

01:06:33.480 –> 01:06:38.740
Ranjit: So we used 3 different Llms now, just just to clarify it.

01:06:38.780 –> 01:07:06.800
Ranjit: The autonomous platform runs inside Google Cloud. So any of the Google assets that we use don’t leave the cloud perimeter we have. So we have a a capability to do some of the work within our own cluster. So it doesn’t even get into Google’s hands. But anytime we want to use some of their bigger models like the text by some 32 K. It goes outside of our cluster into the Google per cloud. But it doesn’t, does not leave the Google parameter.

01:07:07.140 –> 01:07:27.440
Ranjit: So we, we we can decide if the information is needs to be protected or not, and in this particular case we did not. Now, anytime we used openigpt for retail assistant. We are disguising the information, using that homomorphic encryption approach. II describe where

01:07:27.600 –> 01:07:31.929
Ranjit: we’re mapping names to different names. So the actual query that

01:07:31.970 –> 01:07:42.000
Ranjit: the third party engine is seeing is different from the name of the person. And so and and then, when the response comes back, I map it back to the original name of the

01:07:42.080 –> 01:07:46.460
Ranjit: you know of of the customer. So in this case.

01:07:46.550 –> 01:07:49.160
Ranjit: we’re using homomorphic encryption to

01:07:49.490 –> 01:08:06.100
Ranjit: you know, masquerade, the information that goes to the third party engine in this case, since it’s public data. There’s no security in that sense. All the data, exactly as it is typed is sent to the third party engine. So different

01:08:06.130 –> 01:08:08.850
Ranjit: levels of security for different use cases

01:08:09.400 –> 01:08:15.689
Renga: this sort of touches on another point I was talking, I wanna show this slide.

01:08:16.880 –> 01:08:23.290
Ranjit: so we did kind of an an an analysis of you know how you would

01:08:23.310 –> 01:08:24.830
Ranjit: pick an LLM.

01:08:25.670 –> 01:08:46.640
Ranjit: For different types of use case and what what you know what’s the quality and what’s the setup cost and operating costs and things like that? There is a tendency for people to say that. Look, I don’t want my data going to the third party, Llm. Or model. I want to host it in house. So you know, very obvious choices are Llama 2 and cloud 2 which are too powerful open source Llm’s

01:08:46.710 –> 01:08:57.810
Ranjit: which is fine, but it’s important to understand that the cost of running a local Llm. Is pretty high and you’re you’re looking at a few $1,000 per month just to run.

01:08:58.140 –> 01:09:05.899
Ranjit: And Llm. I mean we. I’m not even talking about training your own. Llm. This is taking an open source, Llm. And running it locally with customization.

01:09:06.080 –> 01:09:08.550
Ranjit: The other thing people have discovered is that

01:09:09.060 –> 01:09:24.500
Ranjit: when you bring an open source Llm. In. You’re bringing it with the assumption that it’s risk free. But if there are exploits within that Llm. That people know about those can be exploited. So it’s it’s all you know. The whole security framework is still a little bit

01:09:24.840 –> 01:09:50.439
Ranjit: shaky these days, but I think your best bet is to you know, for ease of use, to get things going immediately. Work with these third party Lms, you keep your costs under control. Understand your use case and then over time. If you decide that you have enough traffic where Roi makes sense for you to host your own Llm. Then you can do that. So you know, that was just a a random side note. Hopefully, it was useful.

01:09:51.680 –> 01:09:53.050
Renga: absolutely.

01:09:53.109 –> 01:10:11.230
Renga: the next question that comes up is from Rabi? He asked. In the context of structured data sources like, you know, my sequel post, Chris, database. Could you give us guidance on how to train the model against the database if it’s not confidential at this point.

01:10:11.480 –> 01:10:24.900
Ranjit: Okay? Yeah. So what we do is we, we have a 2 step process or you know you. Let’s say you have some random structured database against which you want to have a chat chat board.

01:10:25.140 –> 01:10:42.219
Ranjit: So step number one is, we need to sort of understand the schema of that database. And you know, we need to understand to some extent. If there are some obscure names of columns and rows like that, we need to understand what they mean. So step one is to create some kind of

01:10:42.450 –> 01:10:58.729
Ranjit: you can say small prompts. So what we do is so, for example, when I in in my database, when I typed in Joan Kozak. What we do is we actually go into the database. We actually pull some contextual information about this person and create a few different little prompts.

01:10:59.370 –> 01:11:17.450
Ranjit: Okay, so that’s step one. So if so, step one is to understand the database enough to create a prompt so I can create a prompt that, says Joan. Kossack lives in San Jose, California, and has made 16 orders or something like that. So I create a bunch of prompts. And then the second part is, when when anytime I ask a query.

01:11:17.660 –> 01:11:24.510
Ranjit: I need to translate that into the appropriate SQL. And here the genei tools generally

01:11:24.770 –> 01:11:42.870
Ranjit: tend to do a pretty good job, but not all the time. So if I’m doing something like a complex query like where I have to do a a group and distinct or something like that. You! You have to be a little bit careful. So you do. It’s a you assist the prompt with the SQL. Query, and then you can submit that

01:11:43.030 –> 01:11:59.080
Ranjit: to the back end. So I you know I that’s a little bit vague, but I don’t want to give away too much about how we do this. But the but the point is that if you can point us at at any structured database with the schema, we’ll be able to create the prompts and the sequel

01:11:59.330 –> 01:12:05.929
Ranjit: you know, and and have an understanding of the sequel, so that we can run these. So run all these different types of queries.

01:12:09.170 –> 01:12:15.450
Renga: Thanks. Sanjit one last question. And Rabi just commented on

01:12:15.580 –> 01:12:38.910
Renga: the answer to his question, and he said that that definitely helps sanjit one last question from the group. And please keep your questions coming. We have ample time to answer all of them. One last question that we have still open on the chat is, could we speak a bit about the competition as we have covered conversational AI today? What are the customers evaluating

01:12:38.940 –> 01:12:53.879
Ranjit: code AI or yellow.ai as platforms. Yep. So both code and yellow. I, my understanding is that Yellow has actually built their own foundation model. I think they’ve trained their own. Llm, so it’s completely from scratch.

01:12:53.980 –> 01:13:10.809
Ranjit: which is great in one way, because it’s a you know, it embodies all the knowledge that they have, but maybe unproven in a different way in terms of you know. What are the vectors for? Attack and things like that so may you know it. It hasn’t gone through the

01:13:10.960 –> 01:13:28.089
Ranjit: you know. OO Openai! And all of these other open source Llms. Go through a lot of human punching to determine what are the weak points and things like that? So that’s the issue with with the yellow poor, I know has also embraced.

01:13:28.180 –> 01:13:29.209
Ranjit: you know.

01:13:29.460 –> 01:13:39.680
Ranjit: Jenny, I and you know they they’re very sophisticated company and but II think they are using. As far as I know, they’re using third party Llms.

01:13:40.020 –> 01:13:41.830
Ranjit: Now.

01:13:41.880 –> 01:13:48.480
Ranjit: all these guys can probably do something similar to us in some ways, but I think.

01:13:49.210 –> 01:14:16.470
Ranjit: You know. Given that we are looking at at the at the big picture in not just in terms of chat. We’re looking at, you know, using chat as a feeder into back end workflows and things like that. I think we have a different perspective than these guys. And so we’re able to offer a broader end to end kind of solution than yellow or core. But those are both great companies, and you know I’m honored to be to compete with them.

01:14:17.050 –> 01:14:22.650
Ranjit: yeah. So that’s kind of my comment on on Yellow End quote.

01:14:23.850 –> 01:14:38.269
Renga: Thank you. Thank you, Ranjith, and that brings us to the end of the open questions. The QA. Section we definitely have more time to cover for any question that comes up. So are there any questions from the audience

01:14:41.850 –> 01:14:44.940
Renga: going once, going twice.

01:14:46.310 –> 01:15:06.580
Renga: Alright, that’s but thank you once again, Raj, for a very insightful session, I hope. The group goes away with an impression that we are able to build all these use cases using the autonomic platform, and definitely would want to try this out for us, Billy. We also want to impress upon the audience that this is definitely not

01:15:06.910 –> 01:15:19.200
Renga: just a particular webinar occurrence. We tend to have more such sessions like these. And we will definitely introduce more such features and functionality from the platform group.

01:15:19.470 –> 01:15:26.229
Renga: Yeah, I have a response from the question with whom I could not attribute it to. So that’s

01:15:26.340 –> 01:15:41.679
Renga: they said that this session was insightful. Alex asked us if Costa Rica was a hallucination. So it. It was a typo from me. I had copy pasted the wrong

01:15:43.670 –> 01:15:46.960
Ranjit: wrong response. So yeah.

01:15:47.400 –> 01:16:05.979
Renga: awesome. Alex just responded, saying, Thank you. Awesome. Any other questions, any other comments. There’s a flurry of Thank you. Messages. So I’m I’m glad that the great thanks to thanks to everyone who sat through th sat through the whole hour, especially after a holiday. You know any, especially the India attendees.

01:16:06.090 –> 01:16:16.940
Ranjit: I really appreciate your time, and please get in touch with us. If you have any more questions, we’re happy to to work with you, and if you have any specific customer problems, we’re happy to engage with you to solve them

01:16:17.260 –> 01:16:35.270
Renga: absolutely. And just as a footnote everybody was curious to know if they could get a hand on the recording after the webinar concludes. So you definitely get access to the recordings right after the session. And if you’re registered using an email Id, you’ll get a email with the download link.

01:16:35.270 –> 01:16:48.610
Renga: Yeah, but thank you so much for making time, and I hope you see this as one of many more insightful sessions to follow. Have a good day and thank you from myself, Franjith and Mattu from autonomy. Thank you.

01:16:48.630 –> 01:16:52.150
Renga: Yeah. Thanks. Ranga. Thanks. Everyone. Bye.