Feb 25, 2026 Enterprise Brain - V2 - Arch - Transcript 00:00:00 Manisha Gundapuneedi: this thing. I'm not Yeshwanth Reddy Yerraguntla: I think we should pick up the easy ones. Uh but um yeah, whichever way we try to do, we have to come back. Manisha Gundapuneedi: prayer that we have few points. Yeshwanth Reddy Yerraguntla: interview estimate but it's not like we know how much time you're going to take right Manisha Gundapuneedi: Morning. Yeshwanth Reddy Yerraguntla: now like Manisha Gundapuneedi: You have written some points there, right? Like points. Yeshwanth Reddy Yerraguntla: Yeah, eventually bullet Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: points bullet Manisha Gundapuneedi: Yeah, I Yeshwanth Reddy Yerraguntla: points. Manisha Gundapuneedi: think Yeshwanth Reddy Yerraguntla: It'll uh it'll be better. Manisha Gundapuneedi: okay. Yeshwanth Reddy Yerraguntla: What are the epics for each uh epics on coach? What an epic. So tentative guy we are saying we'll be able to do it by uh Manisha Gundapuneedi: for Yeshwanth Reddy Yerraguntla: sixth March but sixth March. Um h yeah tentative guy. Manisha Gundapuneedi: landing. Yeshwanth Reddy Yerraguntla: Yeah, we'll try to say that only 6th March every day 4 hours. Every day whatever we do we have to show up in this document or hopefully somewhere else. 00:03:45 Manisha Gundapuneedi: I'm meeting next Yeshwanth Reddy Yerraguntla: Am Manisha Gundapuneedi: week. Yeshwanth Reddy Yerraguntla: I Manisha Gundapuneedi: Emitting accounts Yeshwanth Reddy Yerraguntla: a meeting? Manisha Gundapuneedi: meeting. Yeshwanth Reddy Yerraguntla: I want to edit custom. Manisha Gundapuneedi: Heaven Yeshwanth Reddy Yerraguntla: Mhm. Sir start anyway because we have to deliver uh to Tata Steel. An Manisha's team will focus as much as possible on uh the unifier agent. Correct. So if we were to say you have to take ownership, Manisha will take ownership of this agent. Manisha Gundapuneedi: Uh she can Yeshwanth Reddy Yerraguntla: Directly Manisha Gundapuneedi: access Yeshwanth Reddy Yerraguntla: see. Manisha Gundapuneedi: Full picture. Yeshwanth Reddy Yerraguntla: Um, backlogs. Backlogs. Yeah. code. No share screen like mainly we are trying to first decide who are the owners for each artifact that we will Manisha Gundapuneedi: Just Yeshwanth Reddy Yerraguntla: deliver. architecture diagram. You have at least uh four architecture diagrams, one for each intelligence agent, right? So four diagrams 00:06:48 Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: and inside architecture diagram what are the things we have to uh deliver as part of diagram and documentation is what we can discuss. could. Manisha Gundapuneedi: We have like enterprise. So first if you again comment down here that will Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: be so we'll have full We can link them. Yeshwanth Reddy Yerraguntla: H. Okay. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: Items obviously taskb 205. Epic Manisha Gundapuneedi: Is there Yeshwanth Reddy Yerraguntla: story Manisha Gundapuneedi: no converter? that. Yeshwanth Reddy Yerraguntla: backlog. So these are things Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: where yeah so first thing is we have to deliver uh uh unifier Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: architecture uh unifier spelling mistake Manisha Gundapuneedi: Agent Yeshwanth Reddy Yerraguntla: What are the other architectures we have to give? Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: Uh core uh agent architecture hint. Manisha Gundapuneedi: Addition to me Yeshwanth Reddy Yerraguntla: for Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: change but work for you must match your project and filters have Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: a status that has assigned to your order. 00:10:24 Manisha Gundapuneedi: Back. Yes, I am descript. Yeshwanth Reddy Yerraguntla: I got a ch Manisha Gundapuneedi: must watch your project and filter and settings. Okay, it's okay. She for now like Yeshwanth Reddy Yerraguntla: 2 minutes. Double work. Manisha Gundapuneedi: okay SP Yeshwanth Reddy Yerraguntla: Hat. Manisha Gundapuneedi: work items from backlog to section. Create new ones. I never doubt. Take a little. I can access it very quickly. And this go Yeshwanth Reddy Yerraguntla: What is Manisha Gundapuneedi: this. Yeshwanth Reddy Yerraguntla: this? Manisha Gundapuneedi: For item, it must match your project and filters and have a status. That's must match your project and filters and have statuses. Start open. TV message. It's a dark Yeshwanth Reddy Yerraguntla: Yeah, I think we spent enough time. Manisha Gundapuneedi: uh sidan can you come to the small meeting room once first office small waiting room that means after 5 10 minutes also That's fine. In Jira, we are creating a task but it's not visible in backlog. 00:17:01 Manisha Gundapuneedi: So that's huh. If we have a goal then we'll get started from there. Yeshwanth Reddy Yerraguntla: Oh, the Excel open Manisha Gundapuneedi: Stranger Yeshwanth Reddy Yerraguntla: Figma document. Manisha Gundapuneedi: when we create a task here it's not visible in backlog create one create Yeshwanth Reddy Yerraguntla: updates Manisha Gundapuneedi: type is Yeshwanth Reddy Yerraguntla: whatever. Manisha Gundapuneedi: finally Yeshwanth Reddy Yerraguntla: Oh, to collaborate with and other teammates in this draft, you need an upgraded SA. Manisha Gundapuneedi: right. So this is not coming in back probably what Yeshwanth Reddy Yerraguntla: Okay. Draft status. Second. Manisha Gundapuneedi: Enterprise and also workflow there is no backlog only. So for example workflow there will not be any backlog not required and then in it has to go to backlog. So whatever the item first you create it will be in the start state only. Start is nothing but backlog only. Okay, it is open. It will go to straight and then at least it should come right. Okay. What are you doing? 00:20:34 Yeshwanth Reddy Yerraguntla: architecture diagram related documents. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: foreign what I'll suggest is um the moment you have access to this uh this link whimsical if you want to make changes you can always open this copy the whole thing create your own whimsical uh dashboard Manisha Gundapuneedi: All right, let's go. Yeshwanth Reddy Yerraguntla: dashboard whatever chart right and paste it. So changes. Manisha Gundapuneedi: Okay. Same like it's same like Yeshwanth Reddy Yerraguntla: So let's focus on this one guy today Manisha Gundapuneedi: We can take like cocoa and cocoa sheet. I'll add like Yeshwanth Reddy Yerraguntla: unify Manisha Gundapuneedi: sheet data sheet. Okay. Yeshwanth Reddy Yerraguntla: Okay. Delete. Manisha Gundapuneedi: Okay, this Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: data source agent will be taken care of. Core agent will be Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: answer them. Um Yeshwanth Reddy Yerraguntla: in the data source agent architecture Manisha Gundapuneedi: oh all of them came in backlog is in backlog only we have to move them to Yeshwanth Reddy Yerraguntla: core agent 00:22:54 Manisha Gundapuneedi: sprint. Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: Okay okay okay. Okay interface. Thank you S. Bye. Done. We cannot quickly Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: part in Yeshwanth Reddy Yerraguntla: agent. I can This is Manisha Gundapuneedi: real time intelligence. Yeshwanth Reddy Yerraguntla: real time decision intelligence agent. Manisha Gundapuneedi: That's Yeshwanth Reddy Yerraguntla: So ownership first thing in the tasks in the deployment architecture Manisha Gundapuneedi: Oh, Yeshwanth Reddy Yerraguntla: and uh uh Manisha Gundapuneedi: just And we Yeshwanth Reddy Yerraguntla: yeah, we'll not go here. Uh Manisha Gundapuneedi: delete. Yeshwanth Reddy Yerraguntla: yeah, basically So decide one of them is clear. do it this way only. Manisha Gundapuneedi: Okay, Sure. Yeshwanth Reddy Yerraguntla: I think will but Manisha Gundapuneedi: Sure. Yeshwanth Reddy Yerraguntla: you unless we architecture I think Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: it's a small thing only Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: as well as list as possible. Uh so first the discuss we have only 45 minutes so what I'm thinking Manisha Gundapuneedi: Okay. 00:26:19 Yeshwanth Reddy Yerraguntla: is uh yeah this is where recordings will also come into picture because we'll talk out loud and uh while talking we'll capture the functional and non-functional requirements. So first of all like let's just discuss a high level understanding. Is everyone okay with what is happening here? Do you want to summarize or uh do you want to poke holes into this uh concept? Manisha Gundapuneedi: So right now we are going to plan. Uh first of all we break into three parts like one is core and another one is Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: data source agent and another one UI. Yeshwanth Reddy Yerraguntla: Got Manisha Gundapuneedi: In core agent we are having two parts right now. Yeshwanth Reddy Yerraguntla: it. Manisha Gundapuneedi: So the one is like real time intelligent plus refresh agent and other one is user insert and other distribution agent. So now we have the owners who will be taking care of what. Now we need to see what are we targeting up front and how we can take Yeshwanth Reddy Yerraguntla: H Manisha Gundapuneedi: them into 00:27:31 Yeshwanth Reddy Yerraguntla: correct. So I'm saying we'll start at this guy data source Manisha Gundapuneedi: parts. Yeshwanth Reddy Yerraguntla: agent. Uh us three will discuss what we need to do. What does it take to get this uh in place? Manisha Gundapuneedi: Right. Yeshwanth Reddy Yerraguntla: So uh first thing uh everyone should agree on is uh there is only one LLM agent here right or uh is there any Correct. Correct. Correct. Manisha Gundapuneedi: Sorry. Yeshwanth Reddy Yerraguntla: Everyone this blue whatever you're seeing is one separate LLM. Yeah. So everyone is a different uh agent. They will not step on to each other. If these two have to communicate that is only through this guy because um like these two need not communicate in any way. This doesn't need any information from this guy right or will there be Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: exceptions? Manisha Gundapuneedi: No. So only whenever we have to take decision or we have to push some escalations, we have to you know agents looking from an outsider point of 00:29:01 Yeshwanth Reddy Yerraguntla: Correct. So the responsibility of data source agent is very Manisha Gundapuneedi: view. Yeshwanth Reddy Yerraguntla: simple. When someone is asking it questions, it has to give back the answers in the most uh efficient manner. It has to be token efficient. It has to be fast. It has to be reliable. It has to be trackable, right? Trackable. And now immediately we'll understand, right? So here you have a connection from this guy to the outside world in some way saying, okay, this needs to talk to some LLM ops platform. Manisha Gundapuneedi: and also intent and intelligence layer. Yeshwanth Reddy Yerraguntla: intented Manisha Gundapuneedi: Whenever it is answering with any uh whenever it is answering any question so it will understand the user intent as well as it will give me the intent response not uh t only Yeshwanth Reddy Yerraguntla: H. Manisha Gundapuneedi: fetching what it is and presenting back or that will be done by uh core agent or something. Yeshwanth Reddy Yerraguntla: Can you like let's take a few examples that way we'll 00:30:18 Manisha Gundapuneedi: So my point is uh what is an Yeshwanth Reddy Yerraguntla: know Manisha Gundapuneedi: escalation is a question from my core agent to this data source agent. So a data source agent should be in a position to understand what it mean to be an escalation. Yeshwanth Reddy Yerraguntla: that will come Here we are feeding it up front what we mean by escalation. Manisha Gundapuneedi: Right. Right. And you should also know like all the sentiment parts all those things also Yeshwanth Reddy Yerraguntla: So Manisha Gundapuneedi: comes into the same picture. Yeshwanth Reddy Yerraguntla: h correct the way it has to store uh it could oh okay I see one crucial thing missing train pattern arc knowledge skills data resources sorry knowledge we'll have that information of uh uh let's say for Gmail agent for every email what is the sentiment associated with it Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: that way some like yeah you have to customize for that agent what is important yeah so if going Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: back to what you're saying which is intent and intention Um if the question is like uh give me all the angry emails from yesterday it can't be fetched from email in any way you have to store that information up front right so if we take a Gmail 00:32:05 Manisha Gundapuneedi: Yes. Yeshwanth Reddy Yerraguntla: example Manisha Gundapuneedi: Element column. Yeshwanth Reddy Yerraguntla: H Manisha Gundapuneedi: So in which while creating embeddings itself it will read all the keys Yeshwanth Reddy Yerraguntla: yeah Manisha Gundapuneedi: how that email is. Yeshwanth Reddy Yerraguntla: correct. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: So if we start uh more details or something schema just a text file sentiment is one important column. Yeah. Manisha Gundapuneedi: Can we think of some examples related to like intent and integers? Yeshwanth Reddy Yerraguntla: Yeah. I got a got a touch examples. Manisha Gundapuneedi: f*** God. Yeshwanth Reddy Yerraguntla: Um, sorry, sorry, sorry. You just want to list down all the examples, right? Manisha Gundapuneedi: Yes. Yeshwanth Reddy Yerraguntla: One example I can think of is give me all angry emails. What else? Manisha Gundapuneedi: Escalations. Give me all escalations or else uh where am I ling in my sales goals? Yeshwanth Reddy Yerraguntla: Hat. Manisha Gundapuneedi: logging in delivery calls also we can prog so in regular and Yeshwanth Reddy Yerraguntla: But the Gmail connect are you uh giving 00:34:30 Manisha Gundapuneedi: not anything yes not only for Yeshwanth Reddy Yerraguntla: examples at a higher level. Okay. Okay. Okay. Got it. H. This is coming from the user directly now. Manisha Gundapuneedi: Yes. I mean initially that may come but system has to the system itself like core agent itself ask these questions to the data source agent on every fetch. All Yeshwanth Reddy Yerraguntla: Okay. Okay. Okay. So I'm understanding there is a concept of uh uh on every fetch what are some what are the interesting uh um things that it has to send back to the main agent. Manisha Gundapuneedi: right. Yeshwanth Reddy Yerraguntla: If every fetch it has to do then it's not like the core agent has to ask them right Manisha Gundapuneedi: Right. Yeshwanth Reddy Yerraguntla: escalations is not optional during uh information needs to be re sent in on ad hoc basis. Yeah. Inside escalations evidence on fetch, right? Manisha Gundapuneedi: Yes. Sorry. Yeshwanth Reddy Yerraguntla: So what that essentially means is um what is the kind of data structure we have to send to dest intelligence we have to decide what that means 00:36:38 Manisha Gundapuneedi: Right. Yeshwanth Reddy Yerraguntla: is data source agent Manisha Gundapuneedi: So, Yeshwanth Reddy Yerraguntla: I don't know. Manisha Gundapuneedi: delivery. Yeshwanth Reddy Yerraguntla: Uh yeah. Manisha Gundapuneedi: So Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: already data source and will take care of Yeshwanth Reddy Yerraguntla: Correct. Manisha Gundapuneedi: and take care. Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: So Yeshwanth Reddy Yerraguntla: Sure. Manisha Gundapuneedi: last half hour Yeshwanth Reddy Yerraguntla: Yeah. Yeah. You can call both Rahul and Arpit. Manisha Gundapuneedi: discussion. Yeshwanth Reddy Yerraguntla: What I'm understanding uh while you're calling them is we have to take those Manisha Gundapuneedi: All Yeshwanth Reddy Yerraguntla: notes. Manisha Gundapuneedi: right. I think it's better if you come with laptop. We're discussing our next steps. Enter. We'll go to that. Uh laptop open Yeshwanth Reddy Yerraguntla: basically for every uh uh what do you call every box we have to understand the data structure interf for the interface. What? How to make these two talk to each other? Manisha Gundapuneedi: Thank you. I send the link if you want to join. 00:39:55 Manisha Gundapuneedi: Thank you. Yeshwanth Reddy Yerraguntla: Here's Jesus. Manisha Gundapuneedi: I'll just Yeshwanth Reddy Yerraguntla: H. Okay. Manisha Gundapuneedi: sh Yeshwanth Reddy Yerraguntla: Yeah. Yeah. This is what we have to do. Manisha Gundapuneedi: Ready? Yeshwanth Reddy Yerraguntla: Jesus. Manisha Gundapuneedi: Listen. Once you are done like I want to ask actually Yeshwanth Reddy Yerraguntla: Yeah, it's not like I'm done in any way. You can ask Manisha Gundapuneedi: so so initially when we started in B like we started with multi- aent architecture like planning agent execution agent like somewhat So later on Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: we like complexity is less and development advancement we went with single agent with multiple tool costs tools basically so right like what will be the architecture here are we Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: definitely sticking with this multi- aent arch orchestration only or it will be hybrid how it works Yeshwanth Reddy Yerraguntla: So the way it's going to work is um whenever you build an agent if the every agent should have the ability Manisha Gundapuneedi: Please. Yeshwanth Reddy Yerraguntla: to uh directly talk to the user interface by itself. 00:43:56 Yeshwanth Reddy Yerraguntla: What I'm trying to say is to attach an agent to pylog is very easy. Manisha Gundapuneedi: Right. Yeshwanth Reddy Yerraguntla: So that way um what you're saying is if all agents are independently functioning of each other but there has to be a common source agent that Pawan started with saying I'll first make a plan and Manisha Gundapuneedi: Thanks. Yeshwanth Reddy Yerraguntla: then um I don't remember how he was uh firing uh individual agents uh but that is what we are going back to because we cannot handle all of these in one with one agent. So we are definitely going to uh stick to multi- aent architecture now. So the pattern is N plus1 agents. Manisha Gundapuneedi: Okay. And one more thing. Yeshwanth Reddy Yerraguntla: Yeah, Manisha Gundapuneedi: So anyway we are completely using the pentic since we started one question ar from the business point of we going to Yeshwanth Reddy Yerraguntla: correct. Manisha Gundapuneedi: clients or someone that question why are we sticking Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: by the definition. Yeshwanth Reddy Yerraguntla: So the reason why we are going with pyic agent is it gives us very low-level control. 00:45:32 Yeshwanth Reddy Yerraguntla: So uh let's say we want to develop Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: something um some concept that has not been done before to do it via pantic AI is Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: far easier and lightweight compared to other agents because every other agent the moment you pip install they come with like hundreds of functions dumped on Not that it's a bad thing but all of those uh what they are trying to do Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: is they try to give you functionality at a uh very high level. If you want to read email it's like they'll say okay there is this function called read email use it. If you want to make any tweaks in that functionality, either you have to go low level and uh uh uh create your Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: own version of read email or you have to subclass that uh read email functionality and tweak it. One of those two you have to do. My uh concern there is uh how many are we Manisha Gundapuneedi: All 00:46:45 Yeshwanth Reddy Yerraguntla: going to do that way, right? because uh agent to agent communication every library implements it in a different way and it's always their own personalized Manisha Gundapuneedi: right. Yeshwanth Reddy Yerraguntla: opinion on how it should work. If I have some problem there and I don't want to use that paradigm but something else I will have to start going deep into the libraries uh functionality and tweak it. Identic AI on the other hand is saying I will not give you anything at a high level. Everything is low level. I will only give you the tools for agent to you make a query it'll give you back Manisha Gundapuneedi: Great. Yeshwanth Reddy Yerraguntla: answer you uh if you want to make it a compatible I have a method um something else something else these low-level abstractions I felt were much better in pantic AI which is why it Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: is we thought I thought it was worth going deep into that only only that library Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: Yeah to some extent yeah we could have gone for any other library also but the level of control pantic gives I think control is equivalent to freedom of uh execution that you can execute in any way you want. 00:48:16 Manisha Gundapuneedi: Okay, got it. Thank you. Yeshwanth Reddy Yerraguntla: Um so I started writing uh what are the things we have to first understand before we say these are the uh tasks right here we are saying uh uh so the here these are all this is called HLD high level design if we start looking into these things we are basically calling it low-level design LLD. So at the least we have to uh put them in Jira saying we have to first come up with these things only then we can execute. Manisha Gundapuneedi: Thank Yeshwanth Reddy Yerraguntla: So there are two types of tasks. Manisha Gundapuneedi: you. Yeshwanth Reddy Yerraguntla: Now for every one of them we are saying first what is the design for these things then based on agreement we can say uh yeah because we everyone agreed on this let's start implementing this. So manga there are two types of uh in things that we have to freeze. One is what are the definitions? What are these trying to do? What are their roles? 00:49:37 Yeshwanth Reddy Yerraguntla: Second thing is uh the uh interfaces. There are two types of interfaces. External and internal. External is like how is this guy talking to outside world. So I yeah we have to come up with more of these things. refresh data evidence. Uh Manisha Gundapuneedi: An one question I'm getting is refresh agent but I mean sorry we put that in the core agent right Yeshwanth Reddy Yerraguntla: no no core agent will simply say give me uh refresh uh refresh your world knowledge and give me the information. Manisha Gundapuneedi: All Yeshwanth Reddy Yerraguntla: But this is the one who is doing the who fetching data and deciding Manisha Gundapuneedi: right. Yeshwanth Reddy Yerraguntla: okay what is worthy of telling it to the real time agent. Real time agent will then consume that and improve build on the world knowledge arena. Manisha Gundapuneedi: Uh coming to data source agent, it will also have its own knowledge also. Yeshwanth Reddy Yerraguntla: That's a good question. In what cases will it need a knowledge graph 00:50:52 Manisha Gundapuneedi: So likewise if data source agent is a Gmail assistant Gmail agent so we are building the knowledge graph for Gmail agent I'm saying what is knowledge graph Yeshwanth Reddy Yerraguntla: H? Manisha Gundapuneedi: means so we are storing all the links of name with its embeddings that embedding Yeshwanth Reddy Yerraguntla: Correct. Manisha Gundapuneedi: is mapped with this link I'm saying that is a knowledge gap. Let me Yeshwanth Reddy Yerraguntla: embedding is mapped with it's not a graph as such it's like you have ID you have the Manisha Gundapuneedi: see. Yeshwanth Reddy Yerraguntla: embedding graph and connection algorithm yeah I mean it's okay that we don't need to go so detailed into it but if your question is uh we have to store some additional knowledge on top of the raw data is is that the Manisha Gundapuneedi: Uh that is what but what is the differentiation between knowledge graph versus this if in knowledge graph is Yeshwanth Reddy Yerraguntla: Oh, this knowledge graph Manisha Gundapuneedi: something it will have a snapshot everything previous state versus new Yeshwanth Reddy Yerraguntla: yourself. Manisha Gundapuneedi: state it's like individual data source agent also will be having Thank you very much. 00:52:13 Yeshwanth Reddy Yerraguntla: Okay. So the core difference between this box this knowledge graph or knowledge base and this knowledge base is this is having this is a snapshot of information across all the silos all the data sources. Manisha Gundapuneedi: And what kind of information Yeshwanth Reddy Yerraguntla: If we take a project for example, we take data steel only that currently in Salesforce it is in so and so lead generation stage. uh in uh Jira uh following epics have been created, following tasks have been created according to emails right like if I open email there will be that uh Gmail summary right this summary according to this summary it understood that uh this is related to enterprise brain or chanaka or Tata steel based on which it will parse the notes and say these are The pending items who is pending on what? No, Manisha Gundapuneedi: Should Yeshwanth Reddy Yerraguntla: this is these pending items are not from Jira, right? This is from Gmail. Then uh you also have probably we covered Jira, email, Salesforce. 00:53:30 Yeshwanth Reddy Yerraguntla: Um from other G we can come right like the Gmail can Manisha Gundapuneedi: I start? Yeshwanth Reddy Yerraguntla: say for Tatastel the task of fetching the connector information from Manisha Gundapuneedi: Mhm. Yeshwanth Reddy Yerraguntla: Tata Steel is pending. Right. So it's not just that I have information of my Gmail. This has information of everybody's Gmail, everybody's Salesforce, everybody's Jira Aranda Manisha Gundapuneedi: So it is always having a state of an Yeshwanth Reddy Yerraguntla: H. Manisha Gundapuneedi: organization Yeshwanth Reddy Yerraguntla: Correct. Project is one concept in the organization. There could be other things also. Manisha Gundapuneedi: even how many sources also there data sources it will be having all of the map like all sources data will be mapped here Yeshwanth Reddy Yerraguntla: Correct. Somehow it has to store all of it. Manisha Gundapuneedi: sources what I did not duplicate Yeshwanth Reddy Yerraguntla: Yes, it is not duplicating. Manisha Gundapuneedi: or Yeshwanth Reddy Yerraguntla: Correct. Okay. Can you elaborate if you have any counterpoint or you want to ask 00:54:38 Manisha Gundapuneedi: each data agent is however having its own knowledge graph. Yeshwanth Reddy Yerraguntla: right here what we are saying is when we talk about or knowledge here we are Manisha Gundapuneedi: Foreign Yeshwanth Reddy Yerraguntla: saying on top of raw data are there any other things worth remembering for example in the email example Rajar is giving there are 500 emails that Diwami has received saved. I need to store for every email what is the sentiment 500 low 400 are neutral 200 are sorry 80 are Manisha Gundapuneedi: speech. Foreign speech. Yeshwanth Reddy Yerraguntla: u 50 or negative 50 or positive that sentiment you don't need to store it here because it's not a snapshot of the organization this is an this is the headache of the Gmail Manisha Gundapuneedi: Okay, cool. Yeshwanth Reddy Yerraguntla: agent because I might ask Gmail Give me all the negative emails that happened from last January to last February. Manisha Gundapuneedi: M Yeshwanth Reddy Yerraguntla: Why will this need to know that? Tell me. Does that make sense? Manisha Gundapuneedi: right. Yeshwanth Reddy Yerraguntla: We are saying we'll store today's snapshot of what is happening in Diwami that seven projects are running in parallel for infinitism. 00:56:03 Yeshwanth Reddy Yerraguntla: these uh sprints have started. Uh these emails have been received from client. Uh we have sent these emails to the client. Uh whatever like for that day it's like the news of that uh project. It is important that you understand why we are storing it. So please ask follow questions. Manisha Gundapuneedi: See it's not like a knowledge graph. Every day the data will get refreshed and it will only have that day's information. Yeshwanth Reddy Yerraguntla: Exactly. It's not like that day's info also like what is relevant to be remembered. It it is it can so happen that I remember something since 1 month in the form of some timelines. It that is not what we are saying here like only today's information. What is important to remember? I'll remember until now some kind of an aggregated uh uh uh information on Kuro. See for Tata Steel there might have been 100 incidents that happened since 2 months. 00:57:19 Yeshwanth Reddy Yerraguntla: We had a lot of email discussions. We had a lot of meetings. Um but is it worth remembering all of them here? I'd probably not. What is worth remembering is what is the latest current progress in every direction for that project. Manisha Gundapuneedi: Thank you. Yeshwanth Reddy Yerraguntla: Jiral nothing has been added yet. That is important to remember. If some things have been added, what are the things immediately necessary to understand? If I say today or today one sprint has uh started uh three people are working on it and uh these are the current tasks on the plate. Tomorrow those tasks might get updated and these three people will take up new tasks. Once a task has been moved from uh you know pending to done maybe after one week I'll get rid of that information from this uh snapshot because I don't need to anymore remember it or even if I want to remember probably I'll remember as a just Manisha Gundapuneedi: What? Yeshwanth Reddy Yerraguntla: a progress saying on this day this task was uh uh this task was attempted and it was finished. 00:58:42 Yeshwanth Reddy Yerraguntla: We don't know exactly maybe that Jira snapshot could be here itself. I don't know. But what what is the purpose of this snapshot is if I ask a question or Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: if this needs to give information to every person in Diwami instead of going here and fetching everything from every agent, I will go here and fetch. That is what is the purpose. Alternative if this is not there. Manisha Gundapuneedi: My question is yes or Yeshwanth Reddy Yerraguntla: Sorry. Uh you want to say Manisha Gundapuneedi: no. So my question is so let's say we remove the direct Yeshwanth Reddy Yerraguntla: something? Manisha Gundapuneedi: dependency on the data source agent whenever a real time information is needed to the user interface we will be relying on the knowledge of the core agent's Yeshwanth Reddy Yerraguntla: H. Manisha Gundapuneedi: uh knowledge right but my question is why we are doing that why we could not rely on the data source agent Why this name? Yeshwanth Reddy Yerraguntla: Because um see there has to be some place of first of all correlating what is happening across different domains right 01:00:09 Manisha Gundapuneedi: Got got I can inter uh I can inter link this way. Yeshwanth Reddy Yerraguntla: Sorry. Manisha Gundapuneedi: So all data source agent gave me like multiple informations. The core agent will interlink multiple data source agents and understand the scenario better. Yeshwanth Reddy Yerraguntla: Ah yeah that is the cross intelligence can only happen here. Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: Okay, Manisha. Uh, did that answer or do you have follow-up questions? Manisha Gundapuneedi: I don't have questions and I'm still Yeshwanth Reddy Yerraguntla: You're still conf Oh, Manisha Gundapuneedi: not 80% I'm convinced but 20% still I Yeshwanth Reddy Yerraguntla: you're still not convinced. Manisha Gundapuneedi: have that question I think it will get resolved when we are discussing that's Yeshwanth Reddy Yerraguntla: Okay. Only thing to keep in mind is in a in a in a in a world where this doesn't exist the burden on getting all information for from every data source will fall on this and that is not the worst case. Once it synthesizes that information, if there are 20 data agents, it will fire 20 requests, get all that information, put all of it in its head and put it on the user interface. 01:01:29 Yeshwanth Reddy Yerraguntla: It that is done. But after half an hour or after 1 hour again it has to go do that whole thing again which is a waste of processing power. Are they if you store last half an hour what happened here? It might say okay just give me whatever happened since last half an hour it will get that info. It will fetch the current state or state half an hour back from here. Then it will know okay these are the things that have changed. First thing it will update this guy because that update is always lightweight. Whatever happened just half an hour back will be hardly anything. So to store will be very lightweight operation. On top of that this guy can always fetch answers from here when the user logs in. This this line is much faster to get answers from compared to this line. Manisha Gundapuneedi: Got it. Yeshwanth Reddy Yerraguntla: But this is always a very heavy LLM. Manisha Gundapuneedi: The why part is good. 01:02:42 Manisha Gundapuneedi: We are we can interpret it. But the how part we are not yet sure how we're going to throw it. Yeshwanth Reddy Yerraguntla: Correct. We have to think through Manisha Gundapuneedi: So we need to imagine like a knowledge graph is a database Yeshwanth Reddy Yerraguntla: that. Manisha Gundapuneedi: having data source with its uh previous state. Data source with its previous state. Likewise, it is mapped or something like an MD file having entire summary with Yeshwanth Reddy Yerraguntla: H it's going Manisha Gundapuneedi: a Yeshwanth Reddy Yerraguntla: to be the best case scenario is to store it by graph. Manisha Gundapuneedi: data. Yeshwanth Reddy Yerraguntla: If you remember Rahul worked on uh graffiti in October November time that will come in handy here. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: So you have to store everything as concepts how the concepts are related to each other and how they are evolving in time. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: You have more doubts yesterday. I have many many doubts actually but I thought good time to ask. Yeshwanth Reddy Yerraguntla: Okay. Whatever doubts you have, what I'll suggest is um put them as part of uh either a Google uh Gmail or in chat. 01:04:04 Manisha Gundapuneedi: Thank you, Yeshwanth Reddy Yerraguntla: Because once you capture them, Manisha Gundapuneedi: sir. Yeshwanth Reddy Yerraguntla: it's always easier to uh then address uh and ensure that they are addressed via some task format or whatever. Manisha Gundapuneedi: Uh yes, Yeshwanth Reddy Yerraguntla: Right. Manisha Gundapuneedi: just to start uh this yesterday we got this Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: excel sheet regarding that uh mapping rules arch types that was much for us not this one like these two are working on simulator. Yeshwanth Reddy Yerraguntla: Oh yeah. Manisha Gundapuneedi: So I to look out for the scenarios look out for 10 Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: to 10 scenarios which we can test over against that framework and Rahul is working on how can we uh read that data all those framing and how can we feed it the best possible Yeshwanth Reddy Yerraguntla: H but I think Ganesh and Anita will be better. Manisha Gundapuneedi: way Yeshwanth Reddy Yerraguntla: No for that because they will work on UI. No. Manisha Gundapuneedi: not we don't needed UI Only we need some simulator which will be having a description. There is no UI for now 01:05:17 Yeshwanth Reddy Yerraguntla: the the task itself is relevant to UI kadant. Manisha Gundapuneedi: for Yeshwanth Reddy Yerraguntla: This is the groundwork for creating goody user interface. Correct? Manisha Gundapuneedi: right. Right. Yeshwanth Reddy Yerraguntla: That is why I'm just saying think through if they are the better person because I feel Rahul and uh Arpit they can work better with the data source Manisha Gundapuneedi: Got it. Yeshwanth Reddy Yerraguntla: agent. Manisha Gundapuneedi: Actually data sourcation these both only we'll be working on now simulator task you are closer to them actually simulator we are Yeshwanth Reddy Yerraguntla: Okay. Okay. Okay. Okay. All I'm saying all I'm saying is there is like Manisha Gundapuneedi: just Yeshwanth Reddy Yerraguntla: how some some work is done by these two some work will be done by those two or not Manisha Gundapuneedi: okay and they already started with Oracle unifi so we let them continue on that for that advantage if Katy and these people will take it forward from today. Yeshwanth Reddy Yerraguntla: yeah Manisha Gundapuneedi: That is why we included them in this 01:06:17 Yeshwanth Reddy Yerraguntla: no here. So coming back to what we are trying to do here. Manisha Gundapuneedi: call Yeshwanth Reddy Yerraguntla: Um someone asked the question and we went here. Um I think Raj Shaker asked. Huh? Sorry. Manisha Gundapuneedi: and knowledge graph. What is the difference? A guarantee that we Yeshwanth Reddy Yerraguntla: H yeah I mean the purpose of these two we I think we Manisha Gundapuneedi: have. Yeshwanth Reddy Yerraguntla: understood but execution part we have to still figure Manisha Gundapuneedi: Yes. Yeshwanth Reddy Yerraguntla: out. Uh so yeah we are almost like at the end of the Manisha Gundapuneedi: Yes. Yeshwanth Reddy Yerraguntla: time this thing what I'll suggest is tomorrow let all 10 of us uh actually meet once because again I want to explain this full architecture to them after which they can take up individual Manisha Gundapuneedi: Okay, Yeshwanth Reddy Yerraguntla: components that's what I'm Manisha Gundapuneedi: then we'll add them to that itself. Yeshwanth Reddy Yerraguntla: saying first half an hour we'll have that after that they can drop off or they can continue so what I'm observing 01:07:22 Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: is um there are only two things to understand at which point we'll know the tasks low level what are the data structures and u what are the interfaces data structures in a sense is like database schema raglo how are we going to store the information um and how are we storing these uh agent uh whatever storage layer alone there that is the data structure I'm talking about data at rest how is it happening for that I I'm calling this definitions but you can as well call it structures and on the other hand we have APIs which are in transportation right when when external agent is asking for some information how should it ask and when it's giving back how should it give back what data structure to give it back in same thing for data source same thing for rag DB same thing for these things. So once you understand these two u I'm not saying we'll be done but we'll Manisha Gundapuneedi: Listen. Yeshwanth Reddy Yerraguntla: have we'll go from 20% clarity to 50% clarity and the last thing that we have to do is uh deployment architecture how are they going to u be deployed right like what libraries we'll use and uh I don't know uh CI/CD. 01:09:21 Yeshwanth Reddy Yerraguntla: How are we going to handle this? How are we going to Manisha Gundapuneedi: communication from agent to this data source Yeshwanth Reddy Yerraguntla: uh Manisha Gundapuneedi: agent. I see Yeshwanth Reddy Yerraguntla: we are mainly talking about the code itself. We also have the diagram. Manisha Gundapuneedi: We can have like a repository setup. Yeshwanth Reddy Yerraguntla: Yeah. And here we'll also have Jira those things. timelines. So because Rahul and Arpit will work on this, I will strongly suggest Manisha uh and you three at Manisha Gundapuneedi: Yep. Yeshwanth Reddy Yerraguntla: least sit for 1 hour today. Manisha Gundapuneedi: Thank Yeshwanth Reddy Yerraguntla: Okay, just to Manisha can explain in her own words what is happening in the whole Manisha Gundapuneedi: you. Yeshwanth Reddy Yerraguntla: architecture followed by why this architecture we are trying to use and uh then the activity will be yeah just to come up with these uh information like check tick marks only so that uh what do you call because you will be implementing you'll have more clarity on uh what we are trying to implement because otherwise if I Manisha Raja discuss that additional overhead of uh us giving you KT will be there right so to to 01:11:30 Manisha Gundapuneedi: Good night. Yeshwanth Reddy Yerraguntla: some extent all I'm saying is uh we'll create owners for these Thanks. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: Rahul is Manisha. Yeah. So, tomorrow we'll we can perhaps continue from here. I'm not saying you have to give an update tomorrow. Uh, all I'm saying is it's better if you can sit for 45 minutes. Manisha Gundapuneedi: So, what are you expecting now? Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: Like, do you want to see something tomorrow like a list of things Yeshwanth Reddy Yerraguntla: I am expecting if not all of these things at least Manisha Gundapuneedi: or Yeshwanth Reddy Yerraguntla: some high level of what are these data structures going to look like? What are these APIs going to look like? Again, I'm not saying there is any proper structure because I myself don't know what is the right structure. Manisha Gundapuneedi: Yeah, we will have to explore. Yeshwanth Reddy Yerraguntla: I think give it a shot. All I'm saying is see how you will explore this, Manisha Gundapuneedi: Sure. 01:12:56 Yeshwanth Reddy Yerraguntla: how you will come up with how will how you will use your own creativity. Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: That said, u I told this already, but uh everyone has this link now. This is a free version. So, Manisha Gundapuneedi: Yes, Yeshwanth Reddy Yerraguntla: it's not like you'll be able to edit this. Uh however, if you want to copy the whole thing, create your own fresh board and simply paste Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: it so that you have your own copy to change. And this you can share or you can present or whatever you want to do. Manisha Gundapuneedi: Okay. Okay. Yeshwanth Reddy Yerraguntla: Okay. Manisha Gundapuneedi: Uh I know we haven't disturbed the original one, right? Whatever we have given to man Yeshwanth Reddy Yerraguntla: Uh but they have snapshots. Why? Manisha Gundapuneedi: because the same link yesterday what I have opened I can see the advancement whatever changes you have added are already added. Yeshwanth Reddy Yerraguntla: Correct. Correct. We are modifying the same link only actually. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: If you want a snapshot, uh, Manisha Gundapuneedi: So, Yeshwanth Reddy Yerraguntla: sure I can do that. I have my own uh this guy, right? So, yeah. Yeah. I mean, this is the snapshot I gave to uh Yeah. So this is there in my personal Gmail. Manisha Gundapuneedi: Okay. Yeshwanth Reddy Yerraguntla: Yeah. Okay. Now so every day we'll have this three and a half of our 4hour Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: discussion. Mainly it's understanding low-level design. Okay. Manisha Gundapuneedi: Okay. Transcription ended after 01:15:11 This editable transcript was computer generated and might contain errors. People can also change the text after it was created.