Mar 3, 2026 Enterprise Brain Tasks - Transcript 00:00:00 Vara Kumar Jagarapu: Okay, thank you. Yeshwanth Reddy Yerraguntla: So I said I'll start with a twominut summary of what you're trying uh enterprise brain Satyasri Prabhakar Mantripragada: Minute. Yeshwanth Reddy Yerraguntla: credentials Vara Kumar Jagarapu: Direct URL. It is login with my Jami. Yeshwanth Reddy Yerraguntla: I can show other things also that's fine actually um essentially The way we have currently built it is uh there is an AI that can get itself attached to multiple sources. It could be Salesforce, it could be Gmail, it could be Google Drive, Google chat, uh any any connector uh and it should be in a position to give these Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: kinds of answers. uh not only answers but somewhat intelligent uh visualizations on the fly, Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: right? Vara Kumar Jagarapu: Yes. Yeshwanth Reddy Yerraguntla: I could have flowcharts on there. Um typical AI kind of a chatbot with uh added twist that it can also do visualizations. Uh that said, Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: the main drawback of this uh whole whatever you're seeing on screen is it's always reactive. 00:01:24 Yeshwanth Reddy Yerraguntla: It will only respond when I say something. We don't want that. We want to have a system that can uh um that can what do you call do things by itself. So one example I can show here Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: is uh content Vara Kumar Jagarapu: Hey Yeshwanth Reddy Yerraguntla: here uh there is a dashboard and if I one thing dashboard by itself should be personalized to every user uh it's like when I wake up tomorrow new item should pop up automatically that that are relevant to me. Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: And uh if I make some if I still the dashboard is still supported by normal chat that if I yeah if I do some uh talking here things will show up here things will get deleted things change whatever very basic uh alpha vision of what we intend to do. There is no absolutely no uh direction in terms of this is what exactly we want to do just a idea on the ch right. Vara Kumar Jagarapu: Thanks. 00:02:43 Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: So it also did some deleting here that said yeah I mean end of the activity we want a proper proactive uh user interface and not a reactive user interface. Dan. Um and on top of that whatever we have built in V1 we realized there are a bunch of um bad practices we went through like we took some shortcuts we came up with the proper architecture. Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: So the ask is to drive this architecture to build this product in such a way that it is as extensible as possible in the sense that End of the day the core will remain as it is. This red box and blue box are red. This is the main agent which is uh doing all the heavy lifting. Blue box is doing the heavy lifting in terms of visualizations and user interface. When a new customer comes and goes, all they have to do is attach their connectors and everything should work. uh uh easily platform if we can build that's that will cover almost 80 to 90% of clients needs. 00:04:10 Vara Kumar Jagarapu: understood. Yeshwanth Reddy Yerraguntla: All right. So to get go a little more deeper on how we want to build it uh pretty Vara Kumar Jagarapu: Mhm. Yeshwanth Reddy Yerraguntla: like as you can see this is the overview you have one two three large boxes to work with. Like I said, this is the user interface uh responsibility. This is the core uh responsibilities. This is the connector responsibilities. We manum data source agents on core agent and user interfaces just for your this thing. I can share. Yeah. Yeah. Satyasri Prabhakar Mantripragada: Enter Yeshwanth Reddy Yerraguntla: Just share just architecture share just Vara Kumar Jagarapu: Oh, Satyasri Prabhakar Mantripragada: and even priority t-shirt size Yeshwanth Reddy Yerraguntla: next. Satyasri Prabhakar Mantripragada: estimates plan. Yeshwanth Reddy Yerraguntla: Yeah, I'm not happy with uh that thing currently that needs a Vara Kumar Jagarapu: heat. Yeshwanth Reddy Yerraguntla: review type. Satyasri Prabhakar Mantripragada: Okay. Yeshwanth Reddy Yerraguntla: So yeah, currently the this is the list of things we have to do now. Um every box uh blue whatever user box core box connector we need user stories we need uh you put HL highle design we need lowle designs also which means we need to understand what are the data structures what are the API interfaces what is this deployment architecture uh what are the logical architecture whatever is on the screen and 00:05:57 Yeshwanth Reddy Yerraguntla: uh to begin with uh uh P kind of an activity for every Vara Kumar Jagarapu: Please. Yeshwanth Reddy Yerraguntla: box um I just started working on the UI uh list of tasks and UI logo I was just listing down what are all the life cycles that are involved based on which what components are being touched to and what can be an MVP. None of these are final. This like I need someone to talk to. I need someone to go through this make updates make changes uh so that more than one people agree on this like where I am sort of struggling is whatever I say yeah they will ask intelligent questions but end of the day whatever they ask I answer they will like accept it correct there's No Satyasri Prabhakar Mantripragada: You need to have somebody who cross questions your uh Yeshwanth Reddy Yerraguntla: challenge. Uhhuh. They do cross question but know sometimes I realize their questions are correct Satyasri Prabhakar Mantripragada: decision. Yeshwanth Reddy Yerraguntla: and I change a diagram here. 00:07:14 Yeshwanth Reddy Yerraguntla: Sometimes uh I give a response and they'll be like okay yeah makes sense right. So yeah and like whatever I'm working on I'm working on here. Main guy if we can come up with some level of MVP that um acts as a proof of concept. Um first thing is we can engage all the juniors for two weeks uh to build these things. how they build, what they build direction, we can let them and meanwhile we can actually sit and work on every small detail in this whole architecture so that once the Jira tasks are presented uh up in after two weeks they can really uh go on a higher gear. So summary Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: Gopal. Yeah. Please. Gopal Gottumukkala: Listening and good so far. Yeshwanth Reddy Yerraguntla: Yeah. Satyasri Prabhakar Mantripragada: Um so just to set the context um now this is the architecture one we need to map it to the vision of the product mostly it would have come that way. 00:08:47 Satyasri Prabhakar Mantripragada: uh if you still feel that in an enterprise context there is some certain aspect is missing or is not fully explored then you may add your inputs uh so that if I if you pitch it no so every customer will look at a different perspective they all should find their answers in the diagram in terms of uh security scalability Vara Kumar Jagarapu: Yeah. Satyasri Prabhakar Mantripragada: reliability uh and Extensibility all aspects whatever Vara Kumar Jagarapu: Uh yes. Yeah. Yeah. Yeah. I'm trying to explore more and so I'll take some time. Satyasri Prabhakar Mantripragada: h Vara Kumar Jagarapu: So most probably like uh yeah initially as uh as concentrating on this user interfaces so that we can discuss once. Yeshwanth Reddy Yerraguntla: Yeah, Vara Kumar Jagarapu: Uh yeah yeah Yeshwanth Reddy Yerraguntla: high level what is expected clarity only I can actually describe the architecture Vara Kumar Jagarapu: yeah Yeshwanth Reddy Yerraguntla: itself. Vara Kumar Jagarapu: what we expected but yeah going to everything is like framework or something that needs to be component right. So 00:10:23 Yeshwanth Reddy Yerraguntla: Yes. Vara Kumar Jagarapu: comp Yeshwanth Reddy Yerraguntla: Correct. Correct. Yeah, I will go through. Shall I go through each box in that case? Vara Kumar Jagarapu: Yes, Yeshwanth Reddy Yerraguntla: Okay. Satyasri Prabhakar Mantripragada: And uh VA just to fill you the the understanding that uh um Vara Kumar Jagarapu: Sharon. Satyasri Prabhakar Mantripragada: we have not finalized any design system or design approach for this application enterprise brain. Vara Kumar Jagarapu: Oh, Satyasri Prabhakar Mantripragada: However, Vara Kumar Jagarapu: fun. Satyasri Prabhakar Mantripragada: design team is working on a research on a a different schema of things that they wanted to implement for AI based applications. How do we present dynamic UI that learns itself and then that uh that uh that presents itself based on the user requirements and user persona. Vara Kumar Jagarapu: H. Satyasri Prabhakar Mantripragada: So our dynamic nature the design team is still researching. Vara Kumar Jagarapu: Okay. Satyasri Prabhakar Mantripragada: So however they have come up with a some certain kind of a framework where they identify each persona that may come to the system to use it and what kind of 00:11:46 Vara Kumar Jagarapu: Okay. Satyasri Prabhakar Mantripragada: information they look for and how do they consume it. Okay. Who comes what do they need? Vara Kumar Jagarapu: Okay. Satyasri Prabhakar Mantripragada: How do they consume that information? Yeshwanth Reddy Yerraguntla: The thing Satyasri Prabhakar Mantripragada: So based on that, for example, if I look for a certain kind of information, Yeshwanth Reddy Yerraguntla: is Satyasri Prabhakar Mantripragada: I may be looking more text based output. You may be looking for it for a graphical based output. Vara Kumar Jagarapu: H. Satyasri Prabhakar Mantripragada: Somebody may need elaborative explanation while somebody need crisp and precise Yeshwanth Reddy Yerraguntla: very small. Satyasri Prabhakar Mantripragada: information. So they have considered all these factors and they have come up with some kind of a framework. So based on that a certain persona when they look for certain kind of information the content that gets generated maybe in one of the four forms for example so that is what they are calling it as a dynamic UI so for each person it changes so that's what we are trying to render it as what was that simulator 00:12:48 Vara Kumar Jagarapu: Okay. Satyasri Prabhakar Mantripragada: UI so anything that Raja or Manisha say we are working on simulator Vara Kumar Jagarapu: Yes. Satyasri Prabhakar Mantripragada: UI the end goal is So to replicate that as of now what Vara Kumar Jagarapu: I understand. Satyasri Prabhakar Mantripragada: we are trying to do is we have uh I think they have done some progress this evening. They tried to show it to me but at that time there was some issue technical issue say they could not Yeshwanth Reddy Yerraguntla: Okay. Satyasri Prabhakar Mantripragada: show. So the UI renders itself based on the persona and Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: Mhm. Satyasri Prabhakar Mantripragada: and so it may still change but the goal is this and we don't want to stick to a chat based interface. Vara Kumar Jagarapu: H. Satyasri Prabhakar Mantripragada: We wanted to break that pattern. The design team wants to break the pattern of providing AI based system as a chatbased interface. chat will be one of the interface mechanisms but not the only mechanism is what they wanted to try it as. 00:13:46 Vara Kumar Jagarapu: Shut Satyasri Prabhakar Mantripragada: Yeah. Yeshwanth Reddy Yerraguntla: Yeah, we'll cover some of those aspects here. Vara Kumar Jagarapu: up. Yeshwanth Reddy Yerraguntla: So first thing first uh when uh user start user has only uh three ways to uh interact uh interact whatever one is the um the UI uh right the main UI Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: UI low as uh Uh is saying UI can itself be u either a dashboard or the uh chat interface Vara Kumar Jagarapu: Yes. Yeshwanth Reddy Yerraguntla: uh sent whatever. So user might say I they might click few things on the dashboard. Uh then they might go to chat interface and say update this task has done or they Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: might say the uh all these five tasks actually assign it to KL. So they might vanish from my dashboard and they might show up on KL's dashboard. Whatever contextual selections based followed by chat interface followed by um there are two personas mainly in our uh whole product Vara Kumar Jagarapu: H. 00:15:19 Yeshwanth Reddy Yerraguntla: end user personas one is uh an end user the other one we are calling as a trainer we'll come to what the trainer means uh but end user personal they might simply give these ad hoc queries that is 90% of what people are doing today right So when I shut down Satyasri Prabhakar Mantripragada: You may use enterprise.com Yeshwanth Reddy Yerraguntla: enter but Satyasri Prabhakar Mantripragada: credential is Diwami all lawyer and password is Diwami Yeshwanth Reddy Yerraguntla: user Satyasri Prabhakar Mantripragada: all lawyer at the at 2026. It's Yeshwanth Reddy Yerraguntla: Okay. 90% of add queries are all chat driven. So you could add your last 121 stage lead count. Wow. Those are ad hoc queries. You get back insights. the uh whatever whatever the core agent needs to deliver it will deliver to the visualization intelligence visualization intelligence decides what are the right Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: way to present it uh and put it on the chart in profession again so uh every time visual this has to 00:16:45 Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: decide what needs to be presented uh connection it is fetching who that user is what are their roles what are their responsibilities, their preferences, everything so that it can customize the same information. So ideally even if co agent is telling exactly the same thing to two different users this is deciding for every user how to customize every intelligence like if you focus on this small icon this is what we Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: I am trying to say as this is the LLM or this is the AI aspect every AI will be attached to certain uh its own Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: database make a Um that database will have information of how it should behave. Skills are on guides on every every one of these boxes will have its Vara Kumar Jagarapu: Thanks. Yeshwanth Reddy Yerraguntla: own uh guidance uh behavior whatever database we'll go through that also. But ad hoc queries key you'll get ad hoc answers. Training lessons on the other hand the trainer is saying uh um trainer is essentially an elevated person in the organization who controls how the whole system can behave. 00:18:09 Yeshwanth Reddy Yerraguntla: So fresh I uh gave this uh whole Vara Kumar Jagarapu: H Yeshwanth Reddy Yerraguntla: interface the whole thing is ready product. They might say these are all 20 processes that run in the company. First is when a project new project starts you have to uh first create Vara Kumar Jagarapu: that's Yeshwanth Reddy Yerraguntla: a uh lead in Salesforce then you have to create a space in Google chat then you have to uh create a project in GitHub then you have to create a share drive send all these necessary links to the um delivery manager delivery owner delivery owner has to be created if they are not created is an escalation to CTO So all that is driven from the chat only by a user. No one is configuring in the back end or there are no configuration files that what any training lessons and that is what we are calling as training lessons and we'll see how Vara Kumar Jagarapu: Huh? Yeshwanth Reddy Yerraguntla: training lessons propagate into the system itself. Vara Kumar Jagarapu: Speech. Speech. Yeshwanth Reddy Yerraguntla: But this is how yeah except for this one part we have we sort of covered how the UI should behave. 00:19:15 Vara Kumar Jagarapu: Speech. Yeshwanth Reddy Yerraguntla: This is interesting in the sense that if the core agent decides something is super critical, it will bypass all of these things and directly send an email or a notification WhatsApp whatever directly to the user. So this is deciding what are important things worthy of people to consume in the organization. This is responsible for customization at a user level. um these two are responsible for sending back some feedback or questions back to the main agent. Coming to the main agent itself, Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: we just covered the user interfaces. Uh yeah, user interface to real time decision intelligence agent work in progress but whatever core agent we sort of covered these four lines. Uh just one second. I just realized user preferences but this is a soft signal. Nobody's sending them like somehow. Uh yeah. Anyway, uh so these are the signals that it receives. Mainly the way this works is there is a we are calling it refresh agent but let's call it refresh mechanism. 00:21:02 Yeshwanth Reddy Yerraguntla: This is pinging uh this agent every minute. there's a schedule associated with it in the in its own database for every data Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: connection we are saying if there are these many agents in our system we need a Gmail agent we need a Google chat we need a Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: Google drive Salesforce Jira min for these five uh uh agents uh every agent will have its own uh frequency of updates. Generally speaking, Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: Jira updates every half an hour. Google chat updates every minute. Gmail updates every minute. Salesforce on the other hand, it updates every 6 hours. This guy is uh saying, "Okay, it's been a minute. Go fetch new information from Google Drive, Google Chat, Google Mail, and Jira. It's been 6 hours. uh go also fetch uh Salesforce if there are any fresh updates. So this is just pinging the main agent saying I think it's time to refresh your own 00:22:08 Vara Kumar Jagarapu: Mhm. Yeshwanth Reddy Yerraguntla: knowledge. So this will ping the data agents uh and the connection has multiple multiple responsibilities on add queries that the user is asking. Sometimes the user is asking super specific question on Salesforce that has to go to the Salesforce agent only. So this will decide I'll ask Salesforce which is an ad hoc query. Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: Um or the same agent can say I want a data refresh. Or the same agent might say here are some training lessons. Please remember them. training lessons uh are like best practices. We'll go what exactly we mean by that when we go to that point. But these are the three things that this guy can ask the data agents. Data agent on the other hand it'll get back with information alerts inside escalation. Essentially all of them are still informations only just that we sometimes it might associate higher importance to some packets of information which based on the packets importance it might decide okay this this is a critical message I'll send it to the user no this is just a normal message uh I will put it in my own database this is somewhat important I have to send it to the user immediately so I'll uh reach this guy so that this guy can send it to the dashboard whatever this it's data 00:23:56 Yeshwanth Reddy Yerraguntla: agents are essentially sending back what they know about their uh data sources so refresh I want to refresh might say uh in the last 1 minute I received 10 emails of the 10 emails uh this itself might decide eight are garbage Vara Kumar Jagarapu: Hey, Yeshwanth Reddy Yerraguntla: I don't even need to send anything but two are actually important I received one email from a customer uh saying they were not happy with something. Vara Kumar Jagarapu: Please. Yeshwanth Reddy Yerraguntla: One or more email could be a new lead in the pipeline. These two will go to the this guy and this guy is going to first importantly update what is called as a knowledge organization snapshot snapshot. Uh let's yeah organization snapshot shot is essentially today what is diwami looking like today what is some organization x looking like in terms of every uh data it got connected to so ikada this is mainly a way of indexing important information so that I can retrieve it uh uh quickly so one way to look at it is there are seven projects running in the company. 00:25:19 Yeshwanth Reddy Yerraguntla: First project uh is almost about to end. Um these are the current emails from customers. These are the current Jira tasks pending. These are the current epics running. Uh blah blah blah. Vara Kumar Jagarapu: Yes. Yeshwanth Reddy Yerraguntla: This is this is the current state. So when anybody comes to the uh user interface and asks what is going on in Infinityism, I don't need to go to every data source agent. I don't need to make all those requests for every agent saying Salesforce low infinum stage entism stage. I don't need to make all those requests. Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: So I'll come here make one request on the snapshot get as much info as possible and get give back to the user that is the sole purpose of organization snapshot. Uh so that's what so when data is refreshed first thing is this core agent is updating the snapshot some uh information is now stale I have to get rid of it some is new information some information I have to overwrite it will take care of all those things. 00:26:32 Yeshwanth Reddy Yerraguntla: Uh second thing is once information is received whatever insights and alerts it decides are important descending to the a different agent Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: called user aware distribution engine where Vara Kumar Jagarapu: Okay. Yeshwanth Reddy Yerraguntla: what is happening is there are these 10 uh snippets of news I have to broadcast to my company but those 10 uh items Not everybody in the company needs to know. Correct. Juniors probably don't need to know. Seniors need to know. But it is out of those 10 items maybe they need to know only two items. Uh but cos based on escalation they might need to know only Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: one or two items. So the core agent's job is to dump what is what is necessary to be broadcast while this agent is deciding for every user how I should convey that information. This is deciding in the 10 items user one needs item 1 3 5 Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: user two needs item 4 7 8 blah blah blah and that is now being sent into the user interface part where visualization intelligence receives for every user uh this is the information I have to broadcast this is deciding how to present it like user on oh he likes bar charts the same information I'll put it in bar chart Right? 00:28:10 Yeshwanth Reddy Yerraguntla: So this is focusing on the how aspect. This is focusing on the what aspect. This on the other hand is like is like a fire hose on just taking everything Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: in and deciding what are all worthy to be sent to the organization. It's not deciding which user what needs to consume. So yeah, I mean the our whole uniquely enterprise brain is Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: in these these three things one 2 three that there is an agent which automatically refreshes its own knowledge. There is an agent which decides for every user what is important and there is an agent which Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: decides how every user will likely to consume the information in the most uh in their own fashions whatever. So snapshot is one critical source of information. User responsibilities we just saw. Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: The third critical source of information for the agent to perform well is the um skills and runbooks. 00:29:21 Yeshwanth Reddy Yerraguntla: This is where the training information now comes into picture. Like I said right when a fresh project starts these are the 10 20 things that happen in my company. Which means when ideally speaking when a when data agent sees a new email for Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: a new project, it might put it in organization snapshot based on this uh skill this knowledge only because runbook is saying uh these are the 10 items. Vara Kumar Jagarapu: Hey. Yeshwanth Reddy Yerraguntla: Uh I will I just received a mail which means I have to put 10 todos on in my uh snapshot. Each todo has to be assigned to some person. Right. Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: So this org info skills and runbooks are all about how the organization is running whereas snapshot is all about what is happening currently. Vara Kumar Jagarapu: H. Yeshwanth Reddy Yerraguntla: The difference between skills and runbooks is runbooks is more of the process itself. Me kitchen example that those are the 10 steps for a fresh project. 00:30:30 Yeshwanth Reddy Yerraguntla: These are the eight steps when there is a bug escalation, Vara Kumar Jagarapu: Yeah. Yeshwanth Reddy Yerraguntla: when there is a problem in production, um when whatever skills on the other hand are like u best practices at a more detailed level. My organization Jira is organized in this format. Salesforce is organized in this format. Vara Kumar Jagarapu: Thanks. Yeshwanth Reddy Yerraguntla: This is how my hierarchy is in my organization. This is not a process in in the sense of steps. This is more like how we what are the thing what what is the most efficient way of going through information. Uh resources on the other hand is like um static documents um things which may or may not change which will again aid runbooks and skills only. is nothing different but resources are sort of static in nature. Guardrails are telling how ARBback is maintained, how uh guard basically safety in terms of AI practices. Vara Kumar Jagarapu: Okay, Yeshwanth Reddy Yerraguntla: Still things are very vague. Only thing I have clarity on is one bucket needs to talk about the snapshot, one bucket needs to talk about organization resources, one bucket needs to talk about user preferences. 00:31:55 Yeshwanth Reddy Yerraguntla: Buckets, all the three are evolving all the time. Vara Kumar Jagarapu: H. Satyasri Prabhakar Mantripragada: Does the guardrails also talk about logs, Yeshwanth Reddy Yerraguntla: Uh Satyasri Prabhakar Mantripragada: logging and auditing mechanisms? Yeshwanth Reddy Yerraguntla: uh AI auditing Satyasri Prabhakar Mantripragada: Um maybe governance or compliance or anything related to that Yeshwanth Reddy Yerraguntla: uh we'll have to decide probably it might sit in skills. Satyasri Prabhakar Mantripragada: logs. Yeshwanth Reddy Yerraguntla: Sorry. Gopal Gottumukkala: You should not let me check. Yeshwanth Reddy Yerraguntla: Yeah, Satyasri Prabhakar Mantripragada: Bluetooth. Gopal Gottumukkala: Give me one. How about Yeshwanth Reddy Yerraguntla: still Gopal Gottumukkala: now? Yeshwanth Reddy Yerraguntla: weak Gopal Gottumukkala: At least it is. Satyasri Prabhakar Mantripragada: It is audible. Yeshwanth Reddy Yerraguntla: audible but very weak otherwise I'll just we'll just increase our Okay, Gopal Gottumukkala: How about now? Okay. To come. So basically uh that should not come here probably that is just a platforms Yeshwanth Reddy Yerraguntla: better. Gopal Gottumukkala: engineering and platform governance these guard rails about because if I would have I Satyasri Prabhakar Mantripragada: H. Gopal Gottumukkala: would be writing my own queries right I would have put in the coping into the picture roles and responsibilities access controls and everything okay so at that point 00:33:26 Satyasri Prabhakar Mantripragada: H. Gopal Gottumukkala: of time we don't even question guardrails because we are coding with them here we are giving LLM access to something and we are saying don't do this don't do this or don't go beyond this those falls into this spectrum nowhere we are Satyasri Prabhakar Mantripragada: H. Gopal Gottumukkala: talking about engineering governance engineering rules and regulations even if we need to talk we should not embed into this diagram it pollutes unnecessarily Satyasri Prabhakar Mantripragada: Okay. Gopal Gottumukkala: so that is what vas What we are looking from is that Vara Kumar Jagarapu: Yeah before going to LLM. Gopal Gottumukkala: also Vara Kumar Jagarapu: So in front of like whatever the information that we need to give we need to decide and need to provide based on Gopal Gottumukkala: yeah that is one aspect and it's much beyond that just as a platform everything that is happening in platform has to be Satyasri Prabhakar Mantripragada: Sure. Gopal Gottumukkala: audited. So when a comes into the picture that could be a little more uh little different and little deeper whatever we are representing here and discussing it is all about A and A's job rather than the engineering aspects of if at all we need to talk about it separate them and then put in a different representation and then finally we'll combine and then show because otherwise it becomes too clumsy. 00:34:57 Satyasri Prabhakar Mantripragada: And uh uh where do we get the token monitoring aspect? Gopal Gottumukkala: token Satyasri Prabhakar Mantripragada: Uh, Gopal Gottumukkala: monitoring Satyasri Prabhakar Mantripragada: I don't not uh Yeshwanth Reddy Yerraguntla: this LLM tokens monitoring a light LLM those Satyasri Prabhakar Mantripragada: uh Gopal Gottumukkala: that falls into the telemetry. Yeshwanth Reddy Yerraguntla: aspects Gopal Gottumukkala: Whatever we usually do will happen here. But here telemetry is captured at a little more uh deeper and differently. Yeshwanth Reddy Yerraguntla: everywhere there is this symbol uh we have to assume three things uh at least three things Satyasri Prabhakar Mantripragada: Okay. Yeshwanth Reddy Yerraguntla: right there will be more one is the actual intelligence itself, the telemetry aspect, uh the telemetry aspect and the logging aspect, right? Um telemetry is more like health, Gopal Gottumukkala: Yeah. Is about Yeshwanth Reddy Yerraguntla: uh response speed, Gopal Gottumukkala: Yeah. Yeshwanth Reddy Yerraguntla: uh how much cost happened. Whereas logging is more about what is the content itself, Gopal Gottumukkala: Yeah. Yeshwanth Reddy Yerraguntla: how did the agent respond, what question did it receive, what answer did it uh answer, what evidence was it using? 00:36:09 Yeshwanth Reddy Yerraguntla: Uh those three have to be bundled together for every agent that there is no doubt in that. Yeah. Satyasri Prabhakar Mantripragada: Okay. Yeshwanth Reddy Yerraguntla: Um ah so this is the core runner that someone is pinging constantly uh what agent what data to refresh. Satyasri Prabhakar Mantripragada: Hold on. Yeshwanth Reddy Yerraguntla: It will ping the data agents. It will get the refreshed data. It will update its own snapshot. At the same time it will also broadcast uh what are the things to uh tell to the organization. broadcast news on coach and this agent uh decides what to be sent to which user uh and this all while updating its own knowledge base of skills run books snapshot user preferences. Finally, what is happening in data agents itself is uh uh in a sense in a sense if I talk about it this orange box and followed by this green box is what we have built on the screen right I'm not trying to super simplify this is only we have done we have touched other aspects also definitely but core what is happening is Someone is asking me question I will make a I'll make a different request to the data source when the question is who all are currently working in infinitism. 00:37:47 Yeshwanth Reddy Yerraguntla: Jira agents uh understanding of who are working in infism is to make a jira Gopal Gottumukkala: Hey, Yeshwanth Reddy Yerraguntla: query um with the right person names. Gopal Gottumukkala: hey, hey. Yeshwanth Reddy Yerraguntla: It will get back the data and it will give back the answer saying these five people are working. uh Salesforce on the other hand it could be different uh in Salesforce terminology it might be who are the sales people who are the pre-sales people associated with that project point is every data source agent even though their functionalities are different some things are very common for first and foremost is the data source that the LLM knows how to make request to the source and get back data This is the u most deterministic aspect in some sense. Two additional things that are that we understood are critical are first of all what is what is the expectation on the data source agent itself in the form of what is it dealing with and it's Jira it is dealing with uh stories epics all those concepts world model in this case could be like um what does Jira story mean what does Jira task mean what is the physical significance of the technical uh concepts right so I can say the concepts are technical uh face of the coin world model is the business uh face of the coin. 00:39:27 Yeshwanth Reddy Yerraguntla: So next thing that will be important is because someone is saying give me news from yesterday. It needs to understand what does it mean by a snapshot. Right? So it needs to know I should not fetch everything from uh the entire history. Um I should fetch only the delta information. I should synthesize it in the right format and give to my uh whoever requested it. followed by some semblance of what is an escalation mean in Gmail. That escalation could mean negative sentiment. In Jira, that escalation could mean a regression of a task from some final stage to a primitive stage. In Salesforce, it could mean a lead has vanished. Right? So uh what that sense of urgency how to convey to the core agent is that escalation rule. So essentially this is all trying to make the agent understand how it should behave. Uh whereas the other side of this thing is its own uh best practices what it's trying to deal with. 00:40:44 Yeshwanth Reddy Yerraguntla: So runbooks same thing runbook is like uh uh if I'm asking for a snapshot of yesterday there might be a limited set of five to 10 queries that it always has to uh uh call prisari to generate all those 10 queries and send it to data source might be a waste of tokens. If we know that it is standard, we'll put it in the form of uh whatever these runbooks are skills. So that skill is basically a packaged uh functionality. So data source agent can simply call that skill and get the information. I will give a example. uh what in Jira what are all the tasks that moved from uh one stage to the other? What are all the tasks that changed their stage stage in the last 6 hours is the question on there is a specific SQL sorry Jira query that will give me back the answer correct that query if it is always somewhat standard where the only thing that changes in the query is the start time and end time the data source agent we can simply say you just figure out the start time end time and plug it into that query. 00:42:08 Yeshwanth Reddy Yerraguntla: and send it to data source. So the number of tokens that the LLM generated here is hardly 10 to 20 tokens in the form of start time and end time maybe because those times are also dependent on the clock. If that skill was not there every time it had to generate that query again and again and again and the bigger risk is not that we are wasting tokens but out of 100 maybe uh 10 times it is making a wrong query. How no matter how good the LLM is instructed how powerful it is the chance that it makes mistakes are very high in the long term. To avoid that we are saying we'll have these things called skills which are repeatable functions uh which the source agent can use again and again. The same thing applies to the core agent also. Uh so that is what we mean by skills as a database runbooks also if there are multiple skills it has to chain together how to do them long or knowledge to some extent when it while it is refreshing there are some things which may be important for it to store uh outside the data source. 00:43:27 Yeshwanth Reddy Yerraguntla: One example I can think is u negative sentiment emails in Google that end of the day email is just email that it's just a bunch of text if I if someone like Navens is asking give me all the mails where people were angry uh in the last 3 months across all projects. The problem is if I now go to the data source and fetch every email and then check for the sentiment, I'm wasting a lot of tokens. On the other hand, if I have a custom field saying for this email, this was the embedding and this was the sentiment approximately then to query negative sentiment will become much faster and more efficient. So that is what we mean by or knowledge. Uh or knowledge is more like table scheas or back guard rails but custom indexes fields for indexing purpose only. We sometimes want few critical signals from every source of information to be indexed. Uh that is what we mean by custom indexes. Our knowledge is like Salesforce. 00:44:47 Yeshwanth Reddy Yerraguntla: This is how ourback works in Salesforce. Guard rails agent. Finally, we are also saying uh data resources in the form of static ones, right? Um there was this project called supply chain management. uh there they the company gave us uh 2,000page PDF and what they said is most of the queries if they are very hard to figure out from common worldly knowledge most likely that answer is going to be present in that 2page PDF Alan TV we are calling as data resources um that is more or less static in nature active inactive I think I was explaining something to Raja when I wrote it. But we can al always have those this resource is active, this resource is inactive kind of a thing. But this DB essentially has it's still a version of this data in the form of a more compressed uh format in the form of indexes so that question answering could be somewhat little faster. So the engineering aspect of every data source agent itself will be a science and an activity because how can we fetch the right kind of information from Jira so that I can quickly give answers 80% of the time those kinds of things it depends on the agent to agent uh it depends on source to source what are the custom indexes that you have to generate for them Um the very last thing that uh I did not cover is the training lessons 00:46:35 Yeshwanth Reddy Yerraguntla: and query same trainer can trainers lessons can percolate back to the source agent that is when runbooks and skills will get updated. Query C on the other hand is more ad hoc basis. If 10 people are asking a very similar question, but that question was not part of skills or runbooks, I still it still makes sense to remember the question and the answer I gave back to the first user so that for the other nine users I can um give back a more faster answer I low um there are those let's say 10 to 20 questions which are generally asked by everyone all the time. Those 20 questions every user will ask in their own language. Right? One guy might ask in passive tone, one by one guy might ask in active tone. Still that doesn't mean there are different questions. So query caching is basically here is the question that someone asked here is the SQL response that uh I made so as to give back the answer. 00:47:49 Yeshwanth Reddy Yerraguntla: Now tomorrow if someone asks a different question but if that question is very similar to this question I am better off just fetching that SQL response from K plugging that SQL to the data source and getting back info. Of course arbback uh security all those things should be taken into consideration at that time. So these are all the moving parts uh uh in enterprise brain in unai. Yeah, mainly I don't know I will continue rambling until you stop me but mainly I am currently like uh creating blocks of everything and for each block I'm seeing what is the life cycle uh associated with that in terms of user experience what components it is touching uh what data packets it's traveling to and if I have to do some MVP just on that one component what would it look like on earth to Rashan? Um but what is the right direction to take this forward because there are so many things we'll have to brainstorm and agree upon. I think one last constraint I want to bring to picture is there is a project called Tata Steel that is coming into pipeline soon. 00:49:25 Yeshwanth Reddy Yerraguntla: India they will be signing India soon at which point we have to deliver the UI part the core agent part and uh couple of uh data connectors for them. It's not finalized which data connector it is but uh we have to deliver at a high level some functional uh prototype at the end of 4 weeks. Um they are not they are very clear that they will not have something at the end of four weeks that they can use at an enterprise level but they said we want to see some functionality at least. Uh yeah 4 weeks is a soft soft 4 weeks is a deadline we have to decide what is the best thing we can deliver by that time. So alpha on coach, beta on coach. Yeah. Gopal Gottumukkala: There's a sales call Yeshwanth Reddy Yerraguntla: Oh, Gopal Gottumukkala: happening. Yeshwanth Reddy Yerraguntla: okay. Yeah, yeah, yeah, yeah. I'm in sales call. Gopal Gottumukkala: Can we catch up with tomorrow? Vara Kumar Jagarapu: H sure no problem. Yeah, sure. Yeah, sure. Yeah, I'll be going through. Thank you so much, Gopal Gottumukkala: the numbers that we will update you Vara Kumar Jagarapu: man. Gopal Gottumukkala: know if we can make sure that you know when everything else is ready just that number the moment we adjust the scope and then we're able Transcription ended after 00:51:31 This editable transcript was computer generated and might contain errors. People can also change the text after it was created.