Feb 26, 2026 Enterprise Brain - V2 - Arch - Transcript 00:00:00 Yeshwanth Reddy Yerraguntla: Hi guys, Arpit Pathak: Learning. Yeshwanth Reddy Yerraguntla: thanks for joining so Kalakonda Harshith Rao: I'm Yeshwanth Reddy Yerraguntla: early. Any updates from yesterday? Am I Abhilash Adunuri: Yes. Yeshwanth Reddy Yerraguntla: audible? Abhilash Adunuri: chang. Yeshwanth Reddy Yerraguntla: Mhm. Okay. So, what is the final uh uh suggestion? Abhilash Adunuri: 2.5 flow uh 2.5 flash is exposing changes 3 pro 3 flash 2.5. Yeshwanth Reddy Yerraguntla: H. So we'll use three flash then. What? Abhilash Adunuri: Yeah, sure. Yeshwanth Reddy Yerraguntla: Which one is the cheaper one? Abhilash Adunuri: Building needles pattern. I don't know. Yeshwanth Reddy Yerraguntla: Choose choose already sir. Choose three pro $12 flash $3. Okay. Three flash 3 pro. What about two flash 2 pro? Abhilash Adunuri: 2.5 Pro three flash. Yeshwanth Reddy Yerraguntla: Okay. Kalakonda Harshith Rao: So as yesterday graphs blinking Yeshwanth Reddy Yerraguntla: Uh-huh. Kalakonda Harshith Rao: and bariz Yeshwanth Reddy Yerraguntla: Okay. Kalakonda Harshith Rao: format. Normal Yeshwanth Reddy Yerraguntla: Okay. Okay. Okay. 00:03:42 Yeshwanth Reddy Yerraguntla: So, Germany three flash is uh $3 per million. 2.5 flash was uh $.5. Yeah. Interesting. What can we do? Continue using FL two three flash then for everything we'll use Abhilash Adunuri: Yeah. So, Yeshwanth Reddy Yerraguntla: uh three flash for Abhilash Adunuri: yeah. Yeshwanth Reddy Yerraguntla: now. people have joined. Anita is here. Okay. Raja Rahul Ganesha is at least for the first uh half an hour I think Ganesh is needed to some extent. Rahul knows what you are doing. So I'll start with a quiz question. Um this is for the juniors especially whatever enterprise brain we are building. Um, is it uh reactive or proactive? Uh, Arpit Pathak: Part should be reactant. Yeshwanth Reddy Yerraguntla: what about others? Abhilash Adunuri: It's very active. Yeshwanth Reddy Yerraguntla: It is reactive. Correct. So first of all, what is the difference between reactive and Kalakonda Harshith Rao: Proactive giving 00:06:11 Yeshwanth Reddy Yerraguntla: proactive? Kalakonda Harshith Rao: advances before we ask it will provide things. Active and giving the correct answer for what we ask. Yeshwanth Reddy Yerraguntla: Correct. Reactive is reactive. Yeah, it is after you ask something it is given. So my I didn't these are not the quiz questions. The quiz question now is I think Arpit can stay silent because he sort of knows the answer. What does it take for us to make enterprise brain re uh from reactive to proactive? Any ideas? Manisha, you can ping the rest of the people. Meanwhile, Anita, Harshett, Amia, Ailash. Abhilash Adunuri: Maybe something like we uh with the current information we need a model or something which can predict the future uh optimizations or performance. Kalakonda Harshith Rao: So my idea is to so if user is asking all about the graphs or Yeshwanth Reddy Yerraguntla: Okay. Kalakonda Harshith Rao: trends over the period. Okay. So we will get the context of the chat. 00:07:47 Kalakonda Harshith Rao: So if we if the next chart or any chart of new chat if the user asks to give me something about this without asking any bar by predicting the uh pattern of user we can make sure he's he's going to ask graph and make it correct like to to know the learn the patterns and make make sure to give the graphs according to the user. Yeshwanth Reddy Yerraguntla: Okay. So, I sense both Abilash's answer and your answer. Kalakonda Harshith Rao: Yes. Yeshwanth Reddy Yerraguntla: What they're trying to say is you try to predict what he's about to ask and then be ready to like answer it up front. Correct. Kalakonda Harshith Rao: As uh many questions were asked between user and LLM conversations, right? So patterns Yeshwanth Reddy Yerraguntla: Right. Correct. Understood. So here also one crucial um um assumption you're making is user will ask uh okay if if a user is asking let's say I am the user and I am interested in understanding things about Tatastel project okay after one week you are saying it will with the architecture you're proposing it will start predicting uh um what I'm about to ask and give the answers up front. 00:09:11 Yeshwanth Reddy Yerraguntla: Then my next question is um what if there is a new project that is going to come into picture? Can it predict that I'm going to ask questions about the new project also? Kalakonda Harshith Rao: We have not mentioned predictions from the customer. Yeshwanth Reddy Yerraguntla: Correct. So that is what we are trying to address with the new architecture. We are trying to address a few uh gaps in our current implementation. where we are saying uh even if we try our best with the current architecture, what's going to happen is can only do incremental updates. On the other hand, in the next half an hour, what we are what I'm going to propose as a system architecture is the ability for the agent to understand everything that is happening in the company whether it is a new project or an old project or a very very old project it doesn't matter right ideally if something is important for me to know I should So if you uh let's say Manisha is the team lead, it doesn't matter whether she asked if people are you know in her squad are on leave or not. 00:11:01 Yeshwanth Reddy Yerraguntla: But as if the agent is really intelligent, it should be in a position to tell Manisha up front that some people are on leave. If it is even more intelligent, it should be able to tell some people are on leave and they have been assigned so and so tasks in Jira. uh given the burn rate we will not be able to do so and so what is the alternative if it is even more intelligent it'll tell the alternatives also assign this this thing or dep prioritize this and uh I'll draft an email for Navins saying this is what is about to happen if it is even even more intelligent it might do those things after approval in Jira only you say okay I'll remove this task from sprint. I'll add this task into from backlog into current sprint. I'll assign it to whatever. So the critical thing that we are missing is um a database where agent needs to know everything about the current system. Today what is the shape of Diwami? 00:12:08 Yeshwanth Reddy Yerraguntla: Tomorrow what is the shape of Diwami? Yesterday what was the shape? Right? If it knows that then only then it can know what to answer what to uh tell up front. So we are proposing an architecture. Uh just one second guys. Sorry. So what you are proposing is an architecture where we don't have single agent. We'll have multiple agents each doing a separate dedicated thing. So if I come to the point where we uh sort of we are currently what we are doing is only this thing we have a data source agent we have a data source you ask something it'll give back answers to a large extent uh this is what we are doing on top of that to some extent we are also doing some intelligent uh visualization manipulation Right based on the question it knows which chart to produce whether it is pie chart or bar chart or whatever and it is answering in that fashion. 00:13:53 Yeshwanth Reddy Yerraguntla: So just before this call uh I was we were working on some manufacturing uh data set and what is happening here it is it is hooked to a CSV file. It is hooked to a uh graphing library. Okay. So based on my questions it is giving decent uh uh visualizations sometimes bar sometimes histogram. But again like I said like we discussed this is not a proactive system. This is a very very reactive system. Nothing is wrong with it being reactive. It's just that people are not impressed. Everybody is building a reactive system. Everybody can build this now. To some extent we also uh increase the complexity of a data source agent by uh adding certain things like uh trained patterns or skills or data resources. Some of you have done this in the project optin where uh it's this is not a rag db but at least the agent was hook hooked to uh online expert it was hooked to run books. 00:15:05 Yeshwanth Reddy Yerraguntla: What is the what are all those things doing or they are trying to give you the best answers for regular questions right? So if if people are asking the same question 100 times uh 99 times it should uh remember that okay this question has already been asked let me store what is the right answer and give it back in some intelligence fashion to some extent. Yeah. So we are we h we we have built a data source agent which was able to connect to data source which is able to do some intelligent uh caching. Of course to some extent we have done this also which is basically we are explaining what the agent itself is doing but what is missing is when we are about to scale the whole system to multiple agents. This is something that Pawan and Abilash and Hershit not hers am Amula were trying to do. At some point uh we were trying to create a planning agent. That planning agent will first come up with the plan. 00:16:12 Yeshwanth Reddy Yerraguntla: Then it will dispatch the right question to the right agent based on which it will get back the answers and it'll try to synthesize. there the problem it was not much of a problem just that debugging speed was slow and uh we I thought it was getting really complex so I paused it and said have a single agent that can talk to every one of them in one shot now what I'm proposing or what Navin and I we discussed and what we are now proposing is let each agent be its own intelligent agent that data source agent is only responsible for giving the best answers for that one data source. It it it doesn't care who is asking the question. This is doing only one thing and that one thing is uh if I if someone asks a query I should give back information. It's an ad hoc uh on an ad hoc basis. Um however if if people are saying uh people are sending in training lessons like you know when you are fetching information from this source you are currently running these kinds of queries but the best way to run it is using this kind of a query don't use select star use something else whatever these training lessons it needs to remember in the form of its own database again to some extent this is already happening in optin in the form of runbooks and that is the 00:17:53 Yeshwanth Reddy Yerraguntla: pattern we understood that is a very valuable pattern for the whole system to learn and grow. Third one and this is the one that makes the whole activity proactive instead of reactive is a data refresh request. Someone is saying uh you know what it's 5:00 since the last time you did some uh data ingesting fetch the latest data from that point. Let's say it uh refreshes every 6 hours. So because it is u right now 5:00 I know that you would have refresh stored the data till 11 in the uh morning. So from 11 to 5:00 p.m. get all the fresh data. So do a request get the data store it in your uh database in whatever uh patterns or format data structure that you think is important. We will define those things but store it so that at every refresh you are only fetching only like limited amount of data necessary for that time span. But critically when you're fetching the data and storing it, you give me back what are all the interesting things that happened in those 6 hours right so last 6 hours may I got these emails uh some of them are negative sentiment some of them are positive sentiment or in Jira these 10 tasks have been moved from in progress to sorry QA to done but actually there is one task which moved from QA to uh in progress. 00:19:35 Yeshwanth Reddy Yerraguntla: So something went wrong. Um that is coming from the Jira agent. There might be escalations. One customer is actually very angry with you. Now who is this customer? Who is this you? In this case, we are not talking, but we are saying if it is a Gmail agent, it will fetch every mail from the organization and send it back to whoever is asking right with what that means is this agent is doing only one thing and that is to give answers based on its data. So if I do a lot of refresh uh activities uh let's say I said Jira you refresh yourself Gmail you refresh yourself Salesforce you refresh yourself Google chat also you refresh sometimes what happens is there will be intelligence across the layers that someone needs to listen to. Jira said this task went from QA to pending or in progress. Google chat somewhere in the chat it is you know some QA person wrote hey this bug is still occurring or this new bug started happening. 00:21:01 Yeshwanth Reddy Yerraguntla: Please look into it. These two are correlated but these two are correlated across different sources. Gmail itself wouldn't have known something was wrong. Google chat wouldn't have known something was wrong. Jira wouldn't have known something was wrong. Or even if they knew what was wrong, they don't have the full picture. That is where we have a hub intelligence which is responsible for doing all of this communication. Every 6 hours it is saying fetch data, fetch data, fetch data, fetch data. Whatever Gmail might be every 5 minutes, Google chat might be five minutes. But Jira on the other hand could be every six hours. That is where the refresh uh activity comes into picture. It should be something that is customizable. Now once this agent has that new information, what it is going to do is it is going to first cor crossorrelate with its own current knowledge. Right? Its own knowledge is in the form of these three silos. 00:22:10 Yeshwanth Reddy Yerraguntla: First of all, what is the current snapshot? Okay, Tata Steel, it is in kickoff phase. Current sprint is running. Uh burn rate is so and so. These are the 10 tasks that are supposed to finish today. Of the 10 tasks, nine have finished. One went back. One went back. Yeah. Sorry. Which means not only did it know that something went wrong from this point, but out of how many that one went wrong, if there are only five that are trying to be delivered or if it is a very urgent task that client is waiting on or whatever based on its snapshot, its understanding, it knows the severity of the information it just received. So it will do two things. First it will update its own snapshot uh for the you know current uh 5 p.m. cadence from 11 a.m. to 5:00 p.m. it updated its own knowledge. 00:23:24 Yeshwanth Reddy Yerraguntla: Second thing is it is realizing there are certain key information that it has to broadcast to um broadcast to different users. So even though it is the same uh task that went into regression, there is a mail that people some client is angry on. It is not necessary that the same information has to go to everybody in the organization. Right? For Pratima madam the same information will go in the form of customer dissatisfaction. For Naven the same information it will go in the form of delivery um delivery labs same will go to recha in the form of uh you know we are not meeting the deadline for u let's say goal sir it will go in the form of uh number of bugs that have been uh happening over the last one week have been increasing every and for rest of the people they don't even need to know because they are not part of Tatast. What that means is the same information agent is this green agent is figuring out I'll have to send to each person in a different way and that information what do you call is coming in the form of user preferences and the current snapshot and it will decide okay for pratima madam I'm going to send this for nav nad I'll send this blah blah blah and that is how we are ensuring that the whole system is actually proactive And it is 00:25:06 Yeshwanth Reddy Yerraguntla: personalized, right? And this user agent anyway whatever information it will get uh it will know how to show the users in the right fashion. For every user there might be preferences, you have best practices, charts, all those things so that your dashboard is always customized. One final thing which is sort of missing is sometimes the agent itself might decide this is so critical that I don't need to call this guy. I will directly send some messages in the form of whatever WhatsApp or email to the right people because this is critical that will directly go to the user it's skipping dashboard and all those things right so this is how you know we are trying to even though the complexity has increased uh this is something that we are unable to avoid and we have to build this kind of a system to really you know tell the world that uh and ourselves that we are doing something truly unique and valuable. So yeah like I wanted everyone to attend primarily because we are going to build this architecture in the next uh few months all 10 of us and uh I think from a workload perspective um the first thing that Manisha's team will handle is the unifier agent which is a data agent which I think the KT happened right so it's a rest based uh service so you'll have to start thinking how 00:26:51 Yeshwanth Reddy Yerraguntla: to build all these components into that one agent whereas Rajshaker's team will look into how to build the core realtime agent along with the auxiliary databases and other things um I should have taken a pause in the middle but uh uh I want people to ask questions Rajashekar G: I'm not prepared. Yeshwanth Reddy Yerraguntla: No questions. It's uh clear. High level. Mannam Sai Anitha: New Yeshwanth Reddy Yerraguntla: Correct. Good question. Mannam Sai Anitha: requments. Yeshwanth Reddy Yerraguntla: Correct. Correct. Correct. Good question. So when a new project comes into picture, first thing that is going to happen is the signal is coming going to come from one of the data agents because it's a refreshbased mechanism. Uh there will be a new email saying uh we had this discussion. There will be a new uh item in Salesforce pipeline saying this is a lead gen. There might be a new chat where two people are discussing this so and so person is interested. 00:28:26 Yeshwanth Reddy Yerraguntla: Whatever it is, somehow one of the sources is going to get uh get that information. Correct. Mannam Sai Anitha: Oh, Yeshwanth Reddy Yerraguntla: One of the Mannam Sai Anitha: it's Oh, Yeshwanth Reddy Yerraguntla: sorry one of the thing that you have to bake Mannam Sai Anitha: in Yeshwanth Reddy Yerraguntla: into every agent is what are what constitutes escalation rules. Sometimes when is there is a new project that is also a key critical information. Now that some agent uh is sending the information that there is a fresh project real time dation intelligence agent what it should do is first of all it has uh this organization information on how the organization should run. This is the snapshot of the organization what is happening in my system. This is the user preferences all those things. But this is talking about how the how Diwami should run when there is a new project. First of all, it should uh we should create a lead in Salesforce. Then we should uh have a personalized uh follow-up messages on Gmail. 00:29:39 Yeshwanth Reddy Yerraguntla: Uh then we should have a dedicated uh Jira project. You should have dedicated uh project manager and product owner for those projects. Do you have a squad lead? How many people are assigned? Where is the highle diagram, low-level diagram? Where are the architecture details? Where are the Jira tasks? All those things are going to sit in the form of a single runbook for that fresh project uh um concept. Then what happens when because this agent realize that there is potentially a fresh project. First of all, it will store it in the snapshot. Second thing is it will correlate with the runbook where it is in the uh according to the snapshot. If Salesforce uh lead is not generated then it should know that it should send the notification to Pratma madam saying one of your teammates should uh generate a lead because uh from this discussion so and so it looks like a project is about to start or if the lead is already generated and product owner product manager are not assigned then the notification should go to Navin Nana saying uh these people are not yet assigned. 00:30:51 Yeshwanth Reddy Yerraguntla: if they are assigned or when they get assigned in the next uh 6 hours then it will say okay Jira tasks are pending that will go to product owner now because it knows who is the product owner this is how on a step-by-step basis automatically it will send notifications to everybody uh by itself Peace. Arpit Pathak: So every note should be categorized in some way such that uh this because uh I mean as you mentioned right pratimm should know that it's a customer dissatisfaction but on the same time I mean so should know that it's delivery fault uh and so so every Yeshwanth Reddy Yerraguntla: Good. Arpit Pathak: notification should have categorization at the Yeshwanth Reddy Yerraguntla: Yeah, I mean the message itself will get customized. Um I'm not clear on what the word categorization means here. There is no category as such. Arpit Pathak: uh categorization in the sense forming the message into a specific format for them. Yeshwanth Reddy Yerraguntla: Yeah. Um that can either come in runbooks or uh in user preferences. 00:32:08 Yeshwanth Reddy Yerraguntla: One of those two is where we might store what is the right format to tell them. Did that answer. Arpit Pathak: Yeah. Yeshwanth Reddy Yerraguntla: Okay. Yeah. I mean end of the day some LLM only has to figure out what is the right message but the guidance should come from these three uh dubbers. Arpit Pathak: Uh also one more thing uh as we are mentioning that uh we are going to recognize the patterns. Uh so what I mean in the Yeshwanth Reddy Yerraguntla: Yeah. Arpit Pathak: longer term uh what how should we think of it? For example, right now we are saying that rag should be the uh thing to answer the knowledge answers. I mean knowledge answers actually. But when we say that recognizing the patterns so we have so much of data feed and database uh right uh and then on the basis of customer or the interaction of the person who is using this thing uh how how should we think of Yeshwanth Reddy Yerraguntla: Um, the Best way to ensure that no matter how big is your database or how many runbooks you have the agent is always able to get that information at a moment's notice is via rag only right that is the only scalable solution. 00:33:54 Yeshwanth Reddy Yerraguntla: So the short answer is everything we want to store as knowledge or something that is static or semi-evolving has to sit in a rag DB. But to talk about patterns itself, how to store, what are the patterns to think in from my understanding these four and these four are what uh make up information like these are the silos in which information can sit. All right. Um so everything that is like a fresh thing that you want to implement in the form of a feature or whatever I feel it should fall into one of these eight buckets and based on that bucket uh you know whether you have to send that information to in the track DB what name space you have to send it in and um during sending we have to decide on what are the right uh uh what are the right uh ways to store these things I yeah even I don't have a perfect answer for that until we start implementing or until we start thinking about those things uh uh I don't think we'll have a 00:35:17 Arpit Pathak: Uh I mean why I asked was because in rag it's not it's not easy to manipulate all these things right once it is fed it is fed. uh we should know that once it is getting updated how should we handle that updated thing uh and it should not revert back to the previous state let's say uh a person said that I want a bar and pie chart later after two days it said no I Yeshwanth Reddy Yerraguntla: All right. Arpit Pathak: want only a histogram so it should never return revert back to all uh that two things and it should always come back and say, "Oh, now this person always want the histogram." So, it should always give histogram. Yeshwanth Reddy Yerraguntla: Okay. I mean that's a very specific example you are Arpit Pathak: Yeah. Yeshwanth Reddy Yerraguntla: giving. Arpit Pathak: Uh because I had no more example that's Yeshwanth Reddy Yerraguntla: No, no. I'm appreciating the question itself that yeah this is something that even I haven't thought through. One thing that should happen is every thing that is interactive should have an active inactive flag. 00:36:34 Yeshwanth Reddy Yerraguntla: uh this is the general way of handling those patterns which uh when a user uh let's say on Facebook they they say okay just deactivate my whole account so what actually happens is you don't delete that row you will say just make that one row in the table inactive so that whatever queries you want to make that you are automatically hitting in only active rows. So in your case when the user switches their preferences or something gets updated first of all you'll make the old preferences inactive and uh add a fresh row. So that is how you don't replace anything in the database. Right? So this is not how you mutate it. Ideally every row should be immutable. The next level of answer I can give is here actually it needs to be truly truly dynamic that every day some things are changing. So uh today the Tatas kickoff like the status of Tata Steel is teams have been decided but Jeratas have not been added. uh but after the 12:00 uh sync what it should know first of all there is a big email uh uh talking about uh a meeting happened for so many hours with these 10 people all the people who are talking about uh architecture diagram now that needs to come and get updated here so that way there I think the best way to have a really dynamic uh data that can evolve is 00:38:19 Yeshwanth Reddy Yerraguntla: through graph only like a graph structure using gravity and the advantage in gravity is also temporal that some naturally some nodes get uh inactive based on time some get deferred um so uh yeah I mean that's like a very high level answer I can give you how we can manage the data dynamic ically but your question was also touching uh a different dimension. You are saying how what patterns to think in Arpit Pathak: I mean in what way? Yeshwanth Reddy Yerraguntla: correct. Arpit Pathak: For example, uh Facebook uses this uh what do you call I forgot the term some analysis they do right? Uh I Yeshwanth Reddy Yerraguntla: Uh Arpit Pathak: mean uh no no uh Yeshwanth Reddy Yerraguntla: profit Rajashekar G: Some kind of digital marketing stuff. Yeshwanth Reddy Yerraguntla: tax. Arpit Pathak: emotion based they do uh so for us what Yeshwanth Reddy Yerraguntla: Okay. Arpit Pathak: would be like that for example uh they predict things Yeshwanth Reddy Yerraguntla: So Arpit Pathak: based on our likes or comments or these Yeshwanth Reddy Yerraguntla: correct. I see. Yeah. 00:39:36 Yeshwanth Reddy Yerraguntla: Uh so how can we send that signal to all the other these things, right? Arpit Pathak: Yeah. Yeah. Yeshwanth Reddy Yerraguntla: How can we send those preferences? Arpit Pathak: Interaction based. Yeah. Yeshwanth Reddy Yerraguntla: Right? I am calling them training lessons. Ideally, training lessons should also that I see should also include and uh preferences. Yeah, I mean based on the user's request, this table has to now update saying user was interested in bar charts or whatever. Uh now there is a difference in uh this thing uh so I think um it has to be updated um that's what so based on the feature that you're trying to implement and based on the life cycle of that feature there is a place for everything. If it is something data specific it has to come and sit here. If it is user specific, it has to commence it here. That kind of a uh address we have we are trying to assign to every uh every packet of information. 00:40:52 Yeshwanth Reddy Yerraguntla: The only question is u u what is the right data structure and when will the agents know uh at what time should uh agent decide something has to be updated or something should not be updated. That's uh in general a very uh important question that we have to answer and uh you know the next one week we are going to sit on this only. Arpit Pathak: I I just remembered the term the sentiment analysis. Yeshwanth Reddy Yerraguntla: Yeah. Okay. Sentiment analysis can locate can be located in multiple locations here also. First of all, if we are talking about this user that is diwami user talking about something uh that can go and sit in user preferences but if the sentiment analysis is from a specific source for example Gmail or Google chat might have sentiment but what's it Salesforce need not have any sentiment right there your organization knowledge right that will have the traditional field called sentiment So you will have to define during the building of that data source agent what are the critical things you want to capture from your uh refresh uh activity because first of all refresh will only do in this case let's say there are 600 new emails I will dump 600 embeddings into my DB that will help for a search activity but that will not help me with questions like uh give me all the angry emails because we know this is a very common uh pattern that we want to capture in Gmail. 00:42:33 Yeshwanth Reddy Yerraguntla: It is worth storing an additional uh embedding like what is the sentiment of the uh email and we'll have to write a special lambda function saying you know there is a tool for this Gmail source agent called get sentiment. This way you have to micromanage all the data source agents. That is where yeah building when when we talk about building a source agent uh the whole activity comes into picture. this mainly deciding what to put in the LLM instructions and what to how to store the uh fresh data because this is anyway sort of easy only right you just make a query you'll get back data Arpit Pathak: So we would label the males. Yeshwanth Reddy Yerraguntla: h on the dimension that we think are important. Arpit Pathak: Yeah. Yeshwanth Reddy Yerraguntla: It sounds like a hack but this is what that that is this is the purpose of indexing uh indexes that if you know every day someone is going Arpit Pathak: I got over Yeshwanth Reddy Yerraguntla: to ask a question about sentiment it's worth storing only sentiment because 00:43:51 Arpit Pathak: a Yeshwanth Reddy Yerraguntla: if you think of the opposite right I don't store anything I will constantly fetch the data on demand then we are wasting a lot of Does that make Arpit Pathak: light actually. Uh we'll see more on Yeshwanth Reddy Yerraguntla: sense? Arpit Pathak: this. Yeshwanth Reddy Yerraguntla: So as an activity what I'll suggest everyone is if people are not asking questions study this diagram and I want everyone to just come up with your own perspectives of which agents are interesting and If I were to give you a choice, mainly the choice is lying in these four boxes, right? Wherever you see this symbol, that is where intelligence lies. So you have visualization intelligence, you have user awareness intelligence, you have realtime core intelligence and data connector intelligence. My suggestion for you as an activity is if I have to give you a choice, which one will you pick and start building? Which one is most interesting to you and which one you have to build? Because this is a very big activity and we need all 10 of us. 00:45:42 Yeshwanth Reddy Yerraguntla: Maybe Raj maybe Manisha you can coordinate and see how it goes. If people don't have preferences, they can we can already give something as a starting point. We can have that discussion in for the next 15 minutes. Rajashekar G: Means uh deciding which one need to give to which Yeshwanth Reddy Yerraguntla: Mhm. Yeah, that's what I mean. Rajashekar G: person. Yeshwanth Reddy Yerraguntla: Just want to hear everyone's perspective on what they want to build or are they okay with anything they want to build because end of the 15 minutes what we can say is take this diagram start building the low-level uh system design start coming up with low-level whatever your understanding is sometimes it's system design sometimes it's low-level uh database design sometimes it's like I just want to capture the use cases for that particular box right lot of things are there you have to create functional requirements nonfunctional requirements data structures APIs then I just realized we have uh user stories actually both of them go hand in hand actually all of these three Right? 00:47:11 Yeshwanth Reddy Yerraguntla: So lot of things are there to do for every box we have to do all of these things which is why I'm saying if we have that discussion you guys can naturally know what to start with this is not like a uh what do you call you are on your own you have to do everything I will get involved in each and every one of them just that I can't do it by myself. Rajashekar G: The core agent part need to get started like in bits and pieces. So in which uh the major point we need to see is what are all the Yeshwanth Reddy Yerraguntla: Mhm. Rajashekar G: boxes we can modularize. Yeshwanth Reddy Yerraguntla: Mhm. Uh can you elaborate because uh Okay. Rajashekar G: I'm thinking Yeshwanth Reddy Yerraguntla: If you're talking about code level modularization, is that what you're saying? Rajashekar G: ah yes uh before jumping into the code level modulation itself we need to think of like how all these can be achieved. Yeshwanth Reddy Yerraguntla: Yeah. Rajashekar G: I mean we need to get some picture like uh if you are calling the refresh agent what we are going to put the logic there which kind of logic needs to be there and how that can be configurable for 00:48:39 Yeshwanth Reddy Yerraguntla: Yeah. Rajashekar G: the each uh data source agent basis Yeshwanth Reddy Yerraguntla: Makes sense. Rajashekar G: And how we are going to build this databases like org information or concepts and user information. Yeshwanth Reddy Yerraguntla: Yeah. Rajashekar G: So how we are going to store and where we need to store for that? Yeshwanth Reddy Yerraguntla: So that is low-level Rajashekar G: Yes, for that part also we can do some kind of uh first Yeshwanth Reddy Yerraguntla: design. Rajashekar G: brainstorming and we want to think about some capabilities. Let's say uh previously whenever we working with the visual defect system we tested our capability of like uh can we uh can we detect the errors on the image that is the first part we think and we executed it in a very Yeshwanth Reddy Yerraguntla: Correct. Rajashekar G: small scale. Yeshwanth Reddy Yerraguntla: All right. Rajashekar G: So in the same way if we have these three kind of informations how we can store them and how we can interpret get them interpreted by an agent we need to execute I think that may give us some more confidence Yeshwanth Reddy Yerraguntla: Got it. 00:50:03 Rajashekar G: that Yeshwanth Reddy Yerraguntla: Got Rajashekar G: let's say there are three kind of informations Yeshwanth Reddy Yerraguntla: it. Rajashekar G: how agent is getting consolidated and presenting it back. Yeshwanth Reddy Yerraguntla: All right. I agree. Rajashekar G: And yesterday when we had a discussion with Namaru also like he said like Oracle Unifier needs to be an MVP and he said those points like what we need to present to them. Yeah. Yeshwanth Reddy Yerraguntla: Yeah. not just I think this is also important for every uh activity. Essentially once these are uh created then we can go here from what I feel Raja and everyone we naturally tried to do something uh couple of times in the form of what are called as simulations right said right we'll talk to whatever ch and get a feel of what is Rajashekar G: Right. Yeshwanth Reddy Yerraguntla: happening So I feel simulations are a good starting point for every one of these blocks because they naturally give us a understanding of uh assuming products assuming feature is done. This will give a natural feel of what is uh what we are trying to build which will naturally help us build functional and non-functional requirements 00:51:56 Rajashekar G: Right Yeshwanth Reddy Yerraguntla: right start these will help Rajashekar G: on. Yeshwanth Reddy Yerraguntla: us uh refine the logical architecture logical architecture is what you're seeing on the screen this is what is called that logical architecture here we came up with some level of detail but are we making any mistakes definitely we are we have a lot of holes like I realized okay I didn't I did not consider authentication at all I did not consider LLM ops this observability at all right so these are the kinds of holes that will automatically get filled once you have more clarity after simulation and these two Rajashekar G: Hey, Yeshwanth Reddy Yerraguntla: So this will naturally become step two. I don't know that ideally speaking everyone wants this guy timelines. Everyone wants timelines. Everyone wants a PC but we have to take it as a little systematically. Every one of these needs its own document so that we have a proper sign off only after which you can be very confident in how to build it. I'm not saying you should not touch this at all. 00:53:31 Yeshwanth Reddy Yerraguntla: Some people might be like this is all too much to understand. Let me just start coding and see how I to get a feel, right? So it's not like you will know exactly what APIs to build unless you start building. So it can go back and forth also rau you can decide on these things. You can decide these and update this because of these two these get updated that uh that is how you can take up the activities. But in general, I have a strong feeling that this is a good starting point followed by this guy after that. Yeah, you have to do other things also for I mean tentatively yesterday I wrote for data source agent uh uh we the owners will be Manisha, Rahul and Manisha anyway will lead the whole data source activity. Um, again I wanted everyone else to talk but somehow I end up only me talking. Rajashekar G: Guys, nobody having any Yeshwanth Reddy Yerraguntla: chap. Yeah, I have this. 00:55:08 Yeshwanth Reddy Yerraguntla: I feel this is interesting. Rajashekar G: questions. Yeshwanth Reddy Yerraguntla: I want to actually do that. It's okay. If all the 10 people want to do one box also, it's okay. It's not. There's no agenda right now. Yeah. Phawhan Saii Gajjalakonda: Uh I actually I was confused that uh how we got there like uh if I say let us say in one project in zero some bug has raised How do our system will identify like when it got synced up? Yeshwanth Reddy Yerraguntla: Okay. Phawhan Saii Gajjalakonda: How it will actually identify that bug and it will how will it consider actually a normal bug and a priority bug? Yeshwanth Reddy Yerraguntla: Mhm. Phawhan Saii Gajjalakonda: How will it decide whether it should be um like uh notified through a WhatsApp mail or email and to whom it should make uh like not to a tech need it has to send or to a manager or delivery manager it also has to send how will decide actually on system decide money Yeshwanth Reddy Yerraguntla: the that will happen in this uh runbook uh concept right. 00:56:17 Yeshwanth Reddy Yerraguntla: So when we are dealing with project burndown or project management or project implementation somewhere we have to tell that uh based on the project deadline uh and the uh story points and burn rate you have to be uh ensuring that everybody knows at least two weeks in advance if something is going wrong that information someone has to decide and put it in the runbook. Then first thing is it knows for a project what is the milestone deadline. Second thing is it knows which epics what is the current epic. Then it will it might ask Jira that you know give me the burn rate give me the uh whatever current progress of the sprint because the data agent is also now informed uh in its own runbooks or skills saying okay this guy is asking me burn rate for current project I will give the burn rate I'll also give is it on track or not so it will give the signal like uh it's not on track that information. Now this agent will consume and it will say um hey in the runbook I know that if something is not on track and deadline is less than two weeks I have to do something I got the information it's not on track but deadline is at least three more weeks to go I'll not do anything or if it is less than 2 weeks then here only we are telling who all to escalate it delivery owner uh delivery owner's manager and 00:58:02 Yeshwanth Reddy Yerraguntla: uh uh whatever sometimes it can so happen that um at an ad hoc level someone might say I am really interested in this project so keep me posted for every small detail ada it might get from the uh preferences for that project ka maybe in the arc concepts that information might be there Navin is interested in this project and he needs to know everything on a daily basis. It will synthesize the runbook. It will synthesize the snapshot. It will synthesize the information from Jira to decide whom all to notify. Is that a sufficient explanation or are there any gaps in the Phawhan Saii Gajjalakonda: I feel basically so even though I mention everything in my run book what I what what should happen and a a bug a big bug got raised which is the triggering point for getting notified. Yeshwanth Reddy Yerraguntla: Right. Oh, you're asking where is the trigger bug got raised? Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: It uh first of all what is the starting point? Starting point is the runbook only. 00:59:18 Yeshwanth Reddy Yerraguntla: When there is a bug, it will go and fetch what to do when there is a bug. Phawhan Saii Gajjalakonda: But basically runbook doesn't know whether bug got uh raised or not right data so small bug is created or Yeshwanth Reddy Yerraguntla: Runbook itself is data source will say pug got Phawhan Saii Gajjalakonda: not. Yeshwanth Reddy Yerraguntla: created. Runbook is all about instruction. It's just a system prompt kind of a thing that says Phawhan Saii Gajjalakonda: So basically this all process starts at data source Yeshwanth Reddy Yerraguntla: when huh data source is the ahuh Phawhan Saii Gajjalakonda: only. Yeshwanth Reddy Yerraguntla: that is the starting point Phawhan Saii Gajjalakonda: So data source when a sync sync has happened it will identify the insights right. Yeshwanth Reddy Yerraguntla: correct Phawhan Saii Gajjalakonda: So some um those in those insights will be tracked Yeshwanth Reddy Yerraguntla: correct Phawhan Saii Gajjalakonda: somewhere and that uh now real time uh ratio ends with the run books will decide like whether to uh take an action or discard that inside is it oh based Yeshwanth Reddy Yerraguntla: This will uh based on the runbook it will yeah yeah yeah based on 01:00:24 Phawhan Saii Gajjalakonda: on Yeshwanth Reddy Yerraguntla: the runbook and the alert uh uh severity it can so happen that uh it might simply say uh okay I'll just store it in the snapshot it's not like anything is super critical. Uh I will not send any notification to anyone but I will I will store it in the snapshot because tomorrow that might come in Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: handy. Phawhan Saii Gajjalakonda: So this way like now our uh snapshots can become a huge. Do you think then our then we can concentrate on every each and every single point like then the is very huge our scope of work or scope of observ Yeshwanth Reddy Yerraguntla: Yeah, Phawhan Saii Gajjalakonda: important. Yeshwanth Reddy Yerraguntla: that's what so we need to strike a balance between what is important and what is throwaway because throwaway I can always fetch from the actual data source on demand. Hey, give me all the angry emails from so and so customer from January to March. That much information should not be stored in this snapshot. Snapshot is always about last one week or last two weeks or last one month. 01:01:43 Yeshwanth Reddy Yerraguntla: So automatically stuff should be pruned out of the snapshot. uh because like you said you can't afford it to be very heavy and as a matter of fact every project uh we deal with Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: or everything that every concept we are dealing with like let's say hiring is happening projects are happening uh events are happening right each one of them will have only one current latest update latest status and maybe five to 10 things that are worth remembering because they might be handy. Everything that happened in the past it should not remember. So there should be an active pruning mechanism. If some bug got raised that got resolved, I should get rid of that uh item from here altogether. Phawhan Saii Gajjalakonda: bit. Yeshwanth Reddy Yerraguntla: Maybe I'll store the resolution for one week that it got resolved. Phawhan Saii Gajjalakonda: Um Yeshwanth Reddy Yerraguntla: Okay, I stored it. But after one week there's no action item on that. So why to store it? Phawhan Saii Gajjalakonda: so our snapshots are uh different from our learnings right and like our agent is whatever our system is learning is not the snapshots. 01:03:00 Yeshwanth Reddy Yerraguntla: Whatever the agent is. No, no, it's not the snapshot. No, no. Phawhan Saii Gajjalakonda: Okay. Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: So our learning where and so we we will have learning data as well right. Yeshwanth Reddy Yerraguntla: Huh? actual learning will happen um here in the form of runbooks in the form of skills in the form of uh data specific skills train patterns let's just call this runbook only these are the ones which u actually are what they uh what the agent is learning these will be semi-static every probably once a week or once a month I'll update something in Phawhan Saii Gajjalakonda: No. Yeshwanth Reddy Yerraguntla: the thing but uh this is not information this is how to do things when I ask a query how to how to run the query should I use selear should Phawhan Saii Gajjalakonda: Heat. Yeshwanth Reddy Yerraguntla: I use a view based query should I use uh a function based query some like IRM might have its own preference like someone already created good list of views uh so in our runbooks we'll say only make view based queries. 01:04:11 Yeshwanth Reddy Yerraguntla: These are the views, right? Adi that is how to perform efficiently. But actual information comes and sits here. Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: What is going on? Phawhan Saii Gajjalakonda: Okay. Okay. Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: And then one one more last question was uh user inside agent Yeshwanth Reddy Yerraguntla: Yeah. Yeah. Phawhan Saii Gajjalakonda: actually where I'm looking was combining both Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: user uh user perspective and then insights alert generation. I feel both both uh two these are two separate things is it or not and Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: I mean generating and maintaining insights this is one thing how to showcase that insights to a person based on the uh person's role etc is one thing I I Yeshwanth Reddy Yerraguntla: I also feel that actually ideally this user preferences directly user interface should fetch. Um let's keep this path. I am also not clear upon exactly. See this realtime agent will say these are the 10 things that are very new to me. 01:05:34 Yeshwanth Reddy Yerraguntla: These are the 10 things that someone in the company needs to know. Who is that person who needs to know? This is not deciding because that is too much of a burden. Phawhan Saii Gajjalakonda: Yes. Yes. Yeshwanth Reddy Yerraguntla: Now Phawhan Saii Gajjalakonda: Again basically uh too much of a job on single agent again it will deprecate it performance. Yeshwanth Reddy Yerraguntla: correct. uh that's what so this will only focus on I want to update my state I want to improve my runbooks I want to uh synthesize information and send it to someone that someone will take that information this this is basically taking let's say 20 one paragraph of info so and so bug has been raised so and so thing has went back into um uh regression um this customer is also was sending not this customer. Bug has been raised. Thing went into correction. One customer is angry. Uh some pipeline in Salesforce is not up to. These are the four things it is pinging. 01:06:42 Yeshwanth Reddy Yerraguntla: Based on these four, it will fetch the relevant users from its table. Uh uh sorry, sorry, sorry. What is it? Huh? It will actually I have to do this. So I can't avoid It has to it is fetching for this information uh is relevant to which all users in the organization. Here somewhere in the as the concept of roles is there roles is basically talking Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: about not only where they are in the organization but which activities they are involved with. I am involved in Tatastel these five projects. Manisha is involved in these three projects. Rasher is involved in these two. How is involved in these five so on and so forth. So for every project or for every entity project and instead of just saying project we can also say hiring this person came into hiring. So it might go and search who are all hiring people. So or it might say this new event people are planning for Ravenclaw who are all actually belonging to Ravenclaw. 01:07:49 Yeshwanth Reddy Yerraguntla: Right? that that is how it is getting all the relevant users not only the users but their roles Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: and their preferences. So for every action item it will know okay this one these five bullet points I have to tell to these 50 people in this 50 sentences Phawhan Saii Gajjalakonda: Yes. Yeshwanth Reddy Yerraguntla: that information will go to UI UI UI agent itself will try to also figure out uh what do you call how is the user interested in learning these Phawhan Saii Gajjalakonda: Actually we are we are combining two things here. Yeshwanth Reddy Yerraguntla: Tell me. Phawhan Saii Gajjalakonda: See one thing I can as a user I can request via chart and then I can get an alert where I didn't request even though I don't I like I'm not requesting Yeshwanth Reddy Yerraguntla: Correct. Phawhan Saii Gajjalakonda: it I can get it from alert. So user aware and inside Yeshwanth Reddy Yerraguntla: Correct. That will not comment a picture. No. No. So I did not add that. 01:08:50 Yeshwanth Reddy Yerraguntla: That's a very valid uh point. Phawhan Saii Gajjalakonda: insights. Yeshwanth Reddy Yerraguntla: ad hoc queries is like a bypass highway right so if I add a new connector from this guy right I am saying I'll just add a new connector to user interface again but that is just giving me the ad hoc responses you have ad hoc queries you have ad hoc responses Phawhan Saii Gajjalakonda: Yes. Yes. Yeshwanth Reddy Yerraguntla: So this is this is completely bypassing the whole Phawhan Saii Gajjalakonda: Oh, Yeshwanth Reddy Yerraguntla: um whole Phawhan Saii Gajjalakonda: yes. Yeshwanth Reddy Yerraguntla: thing Phawhan Saii Gajjalakonda: Yes. But some agent has to answer it, right? Yeshwanth Reddy Yerraguntla: that comes from the data agent. Okay, that can come from real time also. Real time will decide should I ask the data agent because it's not there in my organization snapshot or should I directly fetch from organization Phawhan Saii Gajjalakonda: Oh Yeshwanth Reddy Yerraguntla: snapshot. Phawhan Saii Gajjalakonda: okay Dave uh as per my understandings what I thought was when user asks some question no the agent doesn't goes 01:10:01 Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: through data agent data sources Because once Yeshwanth Reddy Yerraguntla: H Phawhan Saii Gajjalakonda: data is already synced, it will be it will be like uh processed and it will be available as Yeshwanth Reddy Yerraguntla: no but that is assuming you're asking Phawhan Saii Gajjalakonda: snapshots. Yeshwanth Reddy Yerraguntla: something some latest information of them what again that question that I'm asking right give me all the angry emails from this customer since uh in 2025 I will not have Phawhan Saii Gajjalakonda: Oh Yeshwanth Reddy Yerraguntla: it that's Phawhan Saii Gajjalakonda: yes. Yeshwanth Reddy Yerraguntla: Yeah, this highway is important that that I agree. Phawhan Saii Gajjalakonda: Okay. Yeshwanth Reddy Yerraguntla: Ad hoc answers. This is the core chat interface. Phawhan Saii Gajjalakonda: Oh. Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: Oh. uh I have like I got some clarity but still uh there are many things to connect together. Yeshwanth Reddy Yerraguntla: That is why I'm saying all of us have to pick. Phawhan Saii Gajjalakonda: So um what I wanted to do was like wherever we have open that we will do this this this we are not clear we are saying we when 01:11:11 Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: somebody is asking we are saying that this may happen here this may happen here right today we are saying that we should try out the things and check out what are the what we can achieve Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: there for example if I see executing real in a what I achieved there and what are the gaps then I will come back and fix this like fix this architecture again like I'm not able to uh achieve these couple of things in this realtime decision m so I have to uh like insert some other thing I should bring bring some other module achieve this then we can like we are compent on whole uh architecture on Yeshwanth Reddy Yerraguntla: Okay. So what is the action item? That's what I'm trying to understand. We want to retrospect what we have already done and see where are the gaps, right? Phawhan Saii Gajjalakonda: Oh yes and basic. Ah yes and then quickly we have to work on those. We have to test on those gaps. 01:12:34 Yeshwanth Reddy Yerraguntla: testing. We are the people who are building it, right? We should have Phawhan Saii Gajjalakonda: like that. Yeshwanth Reddy Yerraguntla: some Phawhan Saii Gajjalakonda: We had one P like we had one PFC like uh from the uh Google meetings we had to generate some insights. Yeshwanth Reddy Yerraguntla: Right. Phawhan Saii Gajjalakonda: We did one MVP there like long back. Yeshwanth Reddy Yerraguntla: Long back. Oh, Phawhan Saii Gajjalakonda: We did one MVP there like Yeshwanth Reddy Yerraguntla: I know. I know. Phawhan Saii Gajjalakonda: some Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: Real for each engine. Yeshwanth Reddy Yerraguntla: Like if you're saying with this architecture can I get meeting recordings or can I is that what you're saying I want to at least simulate that yeah I can do meeting recordings I can simulate uh ability to update my Jira sprint in one shot through meetings is that like is that the thought process you are going with? Phawhan Saii Gajjalakonda: kind uh similar or Yeshwanth Reddy Yerraguntla: Ah the first step is that only no simulate assuming feature is 01:13:55 Phawhan Saii Gajjalakonda: not. Yeshwanth Reddy Yerraguntla: done but Phawhan Saii Gajjalakonda: Okay. Yeshwanth Reddy Yerraguntla: that is assuming you pick up one part of the whole uh pipeline. You can't pick the entire thing or maybe if you want to you can go for it but it's always helpful to focus on that one part saying this is the role of the module in the whole system and you simulate all the interfaces how do critical messages happen how do ad hoc queries happen u and then you might ask okay can I do for for getting meetings uh summarized and then directly creating Jira uh uh APIs out of it. What does it take? It might give you the entire pipeline but you have to focus mainly on what are the things that are Phawhan Saii Gajjalakonda: Oh, Yeshwanth Reddy Yerraguntla: happening inside realtime decision intelligence as an Phawhan Saii Gajjalakonda: yes. Yeshwanth Reddy Yerraguntla: example. Okay. Yeah, Phawhan Saii Gajjalakonda: Okay. Yeshwanth Reddy Yerraguntla: it might give the full pipeline saying user asked this, this happened, this happened, this happened. 01:15:04 Yeshwanth Reddy Yerraguntla: It will touch every module. Sure you can store that in one big document but what is of actual value will be from your side in those 20 steps there will be five steps talking about realtime decision intelligence those five steps you have to zoom in and bring it into those 15 or 20 steps saying inside decision intelligence the actually these five steps that are happening are actually these 20 steps and uh each step this is the data structure this is the API uh signature Yeah, this this module is talking to that module. This subm module is putting data there. That low level information if you simulate you'll have clear idea what is going. Yeah. Phawhan Saii Gajjalakonda: Understood. Yeshwanth Reddy Yerraguntla: So chapi who wants to take what or do you want some time or I just want to hear from you everyone? Phawhan Saii Gajjalakonda: Oh, I am interested. Yeshwanth Reddy Yerraguntla: Yeah. Phawhan Saii Gajjalakonda: I'm I'm interested in building core agents particularly Yeshwanth Reddy Yerraguntla: Okay. Phawhan Saii Gajjalakonda: uh real time with run books and creating snapshots. 01:16:29 Yeshwanth Reddy Yerraguntla: Okay. I want everyone to talk here. Rahul Mishra: Yeah. Yeah. Sure. Like uh I will be uh moving on to the data sourcing building data source agents. Okay. Yeshwanth Reddy Yerraguntla: Okay. Rahul Mishra: Yeah. Yeshwanth Reddy Yerraguntla: I just noted on What about the rest? Arpit Pathak: Yes, I'll also one more work on data sources and seeing this Yeshwanth Reddy Yerraguntla: Okay. Arpit Pathak: system and all. Yeshwanth Reddy Yerraguntla: Okay. Yeah. I'm just taking running mode. Don't uh assume it's frozen or anything. Yeah. Abhilash Adunuri: I'm interested in building the core agent. Yeshwanth Reddy Yerraguntla: Okay. You want some more time or what? If you want some more time to discuss with Raja or Manisha at least tell that don't be silent. Rajashekar G: I think we need to first of all take up into stories and that would make much more clarity to stories. Yeshwanth Reddy Yerraguntla: Take what Rajashekar G: Stories I mean list of chess there. 01:20:21 Rajashekar G: If you get the list of stories what needs to be done uh Yeshwanth Reddy Yerraguntla: user stories are no but user stories by Rajashekar G: then you Yeshwanth Reddy Yerraguntla: itself are the check department who will do those things also at least they can Rajashekar G: Right. Yeshwanth Reddy Yerraguntla: start right like if pawan is saying I'll do core this thing which means he will start with user stories for that uh component Understood. Rajashekar G: Right. Right. Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: Um me and Rajar we will however uh go with the team's choice only right Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: so we'll be Yeshwanth Reddy Yerraguntla: Yeah. Correct. Correct. M supervisa is data agents. Rajashekar G: Her Manisha Gundapuneedi: raja Yeshwanth Reddy Yerraguntla: Yeah, that is this is the overall emodo can't help much. Amulya Maggidi: I'm interested in tourism. Yeshwanth Reddy Yerraguntla: Okay. Yeah. Core by itself is big one only. Well, I think I will go for it because Manisha said anyway we have to this. This is what Manisha you're saying 01:21:44 Kalakonda Harshith Rao: Data source agent. Yeshwanth Reddy Yerraguntla: right? Rajashekar G: citizen Kalakonda Harshith Rao: Data source Manisha Gundapuneedi: uh part is cor if we have to logically Kalakonda Harshith Rao: agent. Yeshwanth Reddy Yerraguntla: How harsh is Rajashekar G: court. Yeshwanth Reddy Yerraguntla: this? Manisha Gundapuneedi: split between two squads uh we have to go like this only right but Kalakonda Harshith Rao: Okay. Yeshwanth Reddy Yerraguntla: Yeah. True. Manisha Gundapuneedi: we'll be left with the UI part UI Yeshwanth Reddy Yerraguntla: Exactly. Exactly. Manisha Gundapuneedi: agent see guys one thing Yeshwanth Reddy Yerraguntla: No one is touching this guy. Manisha Gundapuneedi: I'm seeing in our team is basically you are only looking UI as UI Rajashekar G: I Manisha Gundapuneedi: here the traditional applications web applications which we have developed Rajashekar G: feel Manisha Gundapuneedi: UI is completely different from what we are developing using AI now AI is running the UI in our application it's not static anymore so it's a very challenging task to generate dynamic interfaces if you are you have already worked on react or any other libraries Yeshwanth Reddy Yerraguntla: Yeah, it'll be very interesting. Manisha Gundapuneedi: It's like it's just plain HTML. 01:22:38 Manisha Gundapuneedi: There is no AI work. I'm not going to improve here or something. Yeshwanth Reddy Yerraguntla: Oh yeah, Manisha Gundapuneedi: Because whoever builds the UI for uh enterprise brain, Yeshwanth Reddy Yerraguntla: definitely. Manisha Gundapuneedi: your experience on UI as well as AI will go exponentially to another level. Why I'm saying is because the way we are looking at a user agent sorry UI agent in enterprise brain is it's I mean as I'm saying it's no more static the way the rendering of the selection of the dynamic widgets that you need to render and the way it has to adjust onto the screens along with the responsiveness and everything is what you need to be expert on. It cannot be developed by a normal developer because I have been in the web development from the last seven to eight years. I'm telling you it's not a normal static web page anymore. So don't look it like that. Yeshwanth Reddy Yerraguntla: Yeah, let me show you one thing that I developed that might give you an idea. First thing is on the UI side based on the question we are asking it is 01:23:52 Manisha Gundapuneedi: all. Yeshwanth Reddy Yerraguntla: sorry just two minutes it's figuring out what to show right like how is it even doing Manisha Gundapuneedi: Yeah. Yeshwanth Reddy Yerraguntla: that there is no here it is there's no histogram or line chart component there's no mermaid component right that's but somehow we are still able to achieve this on the fly even showing this was people are not impressed because they are saying anybody can do But this itself is we took so much time. On top of that we built something in the form of a CRUD for UI. CRUD is like create, rip, those things, right? Create, read, update, delete. So through UI we are able to uh basically generate those HTML elements on the fly and update those HTML elements. Right here I'm saying add some dummy approvals. Somehow it it knows what to do. Not only it knows it it should be in a position to do it aesthetically. It should have the context of the UI. It should handle multiple pages in the UI. 01:25:04 Yeshwanth Reddy Yerraguntla: When I say add something here, it should know what that here means. So so much is un uh unexplored in even the state-of-the-art uh communities. If this was possible, you know, Google or Microsoft would have released these kinds of things, right? But this is so fresh and so new that we need we learn so much basically just by attempting this thing and I'm sure we'll do some we will do a very good job. So don't like Manisha said don't assume this is uh conventional in any way. You will apply conventional learnings but you will build on top of it. You'll build something original like every component component is going to be original. This is like a different beast altogether. Manisha Gundapuneedi: So who is taking it up? Rajashekar G: Everyone is silent. Everyone is Manisha Gundapuneedi: H. Rajashekar G: silent. Yeshwanth Reddy Yerraguntla: Shall I Ganesh Danuri: Then manishaka will take this up. Yeshwanth Reddy Yerraguntla: nominate Manishaka will do other things. She has more headaches. 01:27:03 Ganesh Danuri: Just joking. Yeshwanth Reddy Yerraguntla: Unfortunately play mak she would want to do it. She jump onto this opportunity. Unfortunately she has a problem. The problem is she's a senior. She can always closely follow. Shall I shall I suggest uh a solution then? If no one is coming forward I want to nominate two Manisha Gundapuneedi: Anna, Yeshwanth Reddy Yerraguntla: people. Manisha Gundapuneedi: please nominate only from one squad because again we can't split this between two Yeshwanth Reddy Yerraguntla: Yeah, already um I have a feeling that data source agent can be handled by just two people. Manisha Gundapuneedi: squads. Yeshwanth Reddy Yerraguntla: That's sufficient. On top of that, Ganesh I know is very good in UI. He's like smart and he can come up with really original ideas. So I want Syanita and Ganesh to take a PY. That will ensure it's in the same squad. Mannam Sai Anitha: Okay. I know. Yeshwanth Reddy Yerraguntla: Sure. Manisha Gundapuneedi: Ganesha Yeshwanth Reddy Yerraguntla: Yeah. 01:28:17 Ganesh Danuri: and uh I'll guide Anita but uh I also want to work on the core agent actually but it's two different squad seats I Manisha Gundapuneedi: Cora. Ganesh Danuri: guess. Manisha Gundapuneedi: Yeah guys, Ganesh Danuri: Yeah. Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: so I mean one request that we cannot track. Ganesh Danuri: that we can Manisha Gundapuneedi: Awesome. Yeshwanth Reddy Yerraguntla: Echo. Manisha Gundapuneedi: Uh so basically we have to have a logical split between two squads so that we are Ganesh Danuri: try Manisha Gundapuneedi: not I mean fighting on the same thing. So that is why that one compromise we have to go with for sure if squad is going with co core agent the other squad it's requested that we don't touch that Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: part it will be complete responsibility of that squad to complete the it applies same as Ganesh Danuri: to Yeah. Yeah. Yeshwanth Reddy Yerraguntla: Heat. Ganesh Danuri: Uh but I think spending two guys on UA agent is not I mean not required. I guess one guy will be fine. Manisha Gundapuneedi: I'm saying this is not a traditional thing. 01:29:18 Ganesh Danuri: I'm saying this is you need a person Manisha Gundapuneedi: You need a person who to work on the AI. Visual intelligence is the AI layer here which decides on uh what needs to be generated on the UI, how it will fit and everything. Here there is still AI. It's not just UI components that we are writing. We have to first build UI library first with all the widgets and components. We have to work on Yeshwanth Reddy Yerraguntla: Yeah, I mean Rasher can also pitch in and tell how different it Manisha Gundapuneedi: us. Rajashekar G: So currently we are rendering Yeshwanth Reddy Yerraguntla: is. Rajashekar G: HDMX we using like HTMX for serverside rendering and rendering the visualizations based on the question itself the agent itself taking the decision what to render on the UI. This is the major part here. Ganesh Danuri: Uh yeah in in the enterprise brain application what I saw Rajashekar G: But Ganesh Danuri: was everything was generated on the server it was just sending it to the uh front end actually front end is just rendering that uh UI but this 01:30:21 Rajashekar G: client but this one Yeshwanth Reddy Yerraguntla: Correct. Rajashekar G: doesn't complete the overall requirement in future. Why? Yeshwanth Reddy Yerraguntla: This is bad in a lot of Rajashekar G: Because uh this UI is not good feedback every Yeshwanth Reddy Yerraguntla: ways. Rajashekar G: time. So people are looking at like what are the more visualization packages like there is a package called flourish likewise there are multiple packages will be there. Yeshwanth Reddy Yerraguntla: Yeah. Rajashekar G: So we need to look for uh so best packages and how can we integrate it to the best uh client rendering technology like if you go with react there will be much compatibility with this more kind of uh animations and stuff. But if we go with uh this uh plotly or something we end up doing like this kind of uh U. Yeshwanth Reddy Yerraguntla: This is Yeah, Rajashekar G: So basically we will we will start with Yeshwanth Reddy Yerraguntla: sorry. Cut it. Rajashekar G: this if possible in the meantime we will find it like in an iframe we will be able to render ReactJS or something we'll go with that. 01:31:31 Rajashekar G: If that is also not possible if we can pro crack crack that react itself we will directly jump to the react. So once we are jumping into react we need to build everything from scratch again like uh chart interface dashboards every dynamic component needs to be in the front end everything needs to be there. Yeshwanth Reddy Yerraguntla: So there are already some libraries which are doing very similar things right. So these are all pure react uh uh front end uh kind of a thing. So we don't know how this technology works. What we have done here is in the interest of time in the interest of development activities. I I personally pushed everyone to take this route because uh because we wanted to show some progress but this clearly is a very hacky solution that you are you're making AI generate everything then sending the iframe to the front end front end is just pushing that I frame there's no intelligence right if I if I want to do something like I select this guy and then ask a follow-up question I can't do So there is a lot of scope in just cracking the technology itself. 01:33:10 Yeshwanth Reddy Yerraguntla: Yeah, I mean yeah I'm extremely happy if uh Ganesh is proven right that this is so simple that one person is necessary but at least you can start with two people if you can prove that yeah this is not a big deal we'll think of a way to see what can happen but I'm reasonably sure that this Ganesh Danuri: I'm not sure other Yeshwanth Reddy Yerraguntla: is a complex task And Ganesh Danuri: and Yeshwanth Reddy Yerraguntla: anyway like we are saying right this involves AI anyway. So if you can perfect the AI part of this here a Ganesh Danuri: anywhere. Yeshwanth Reddy Yerraguntla: lot of things can be anyway common in here also. So I think u shall we freeze this table? We are already like half an hour late everyone. Okay. Rajashekar G: Yeah. Manisha Gundapuneedi: Can students supporting Ganesh Danuri: Uh actually I Yeshwanth Reddy Yerraguntla: Yeah, Manisha Gundapuneedi: action? Ganesh Danuri: want to work with agents but I'll see we'll Yeshwanth Reddy Yerraguntla: there is an agent. Ganesh Danuri: go. Okay. We'll start. 01:34:33 Yeshwanth Reddy Yerraguntla: only agent. Ganesh Danuri: We'll start. Yeshwanth Reddy Yerraguntla: Yeah. Get started. Okay. Nine out of 10 people are okay with it. Ganesh Danuri: out of 10 people are okay with it. So yeah, Yeshwanth Reddy Yerraguntla: So it's unfortunately. Ganesh Danuri: we'll get started with it after that. Yeah, you'll see. Yeshwanth Reddy Yerraguntla: Yeah. Second uh so next activity is for all of you will be uh take this link create a fresh copy on your create a fresh account uh create uh whimsical fresh account. They'll give you three boards for free. Copy this whole thing and start working on your Ganesh Danuri: Uh and one more thing on uh deploying this application Yeshwanth Reddy Yerraguntla: component. Ganesh Danuri: uh actually I worked with the DevOps team actually uh to how to deploy uh front end application back end application using easy to everything. Yeshwanth Reddy Yerraguntla: Mhm. Ganesh Danuri: So at least for day one and uh QA I'll start deploying this in instead of going to the QA QA guys actually I mean sorry DeOPS guys so uh I can take that 01:35:39 Yeshwanth Reddy Yerraguntla: Okay. Ganesh Danuri: also with the help of them actually yeah Yeshwanth Reddy Yerraguntla: Oh, okay. The deployment aspect. Ganesh Danuri: deployment aspect LLM ops is there right so with with the help of them actually I'll start working on Yeshwanth Reddy Yerraguntla: Right. Right. Right. Ganesh Danuri: that so instead of we going to devops we We we can have that. Yeshwanth Reddy Yerraguntla: Okay, this makes sense to me. Ganesh Danuri: This makes sense. Yeshwanth Reddy Yerraguntla: I completely forgot about this. Yeah, great. I think in from your side it will be like what is the deployment architecture going to look like? All those things. Ganesh Danuri: Uh, Ian. Yeshwanth Reddy Yerraguntla: Yeah. Ganesh Danuri: So I'll work on the UA agent and the deployment Yeshwanth Reddy Yerraguntla: Uh I'll leave it to Manisha to uh collab sorry collaborate with yeah see how we can handle this definitely LLM sorry deployment part I completely missed Manisha Gundapuneedi: Okay, I will check that also. The can you just add that in the table to both me and Ganesh? 01:36:53 Manisha Gundapuneedi: We'll look into that. Deployment parts plus UI plus Yeshwanth Reddy Yerraguntla: Sorry. You have three responsibilities. Manisha Gundapuneedi: deployment. Yeshwanth Reddy Yerraguntla: Okay. Yeah. Yeah. But let's work together on all of these things, guys. This is uh uh working from ground up. We never did it. It was always ad hoc. So go very systematically and achieve this. It'll be good for all of us. So what will be the next steps? Um I'm mainly expecting everyone of us to I don't Manisha Gundapuneedi: I'm to me of blessing. Yeshwanth Reddy Yerraguntla: know that by the way I know most of you don't know this not that everyone of you are needed also but I'll just let you know something that every day we have uh my god stop we have uh whatever four and a half of hours with a break uh just for this discussion for the next one week. So what we'll do is uh whenever someone is ready with some updates they can join in give their updates and leave the meeting. 01:38:29 Yeshwanth Reddy Yerraguntla: Manisha Raja and myself we three will be common everyone else can come in and leave can is might give you the impression that it's optional but everyone should do that every day so that we know what is the progress like so what I will uh ask is uh on Monday everyone try to have u uh these very um minimum on top of that if you have other things also that will be great or what do you suggest I'm bad at planning so I leave Manisha and Rashik to take better questions Rajashekar G: Either way it is good enough like you suggested two ways right Yeshwanth Reddy Yerraguntla: Um. Rajashekar G: we can connect in the morning as usual uh already like uh every day 8:30 we are connecting with everyone right so by that time if everyone one is Yeshwanth Reddy Yerraguntla: Okay. Rajashekar G: ready with their points in like uh their area. They also come there and update Yeshwanth Reddy Yerraguntla: actually the best. Manisha Gundapuneedi: And I would suggest uh sorry I would suggest we'll ask all these people to join 01:39:55 Rajashekar G: us. Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: in the last half an hour so that by the time we we are also done with our discussion we can give them key points and also they can also come up with their status instead of all of us joining because um I mean for us we wanted Yeshwanth Reddy Yerraguntla: Yeah. Manisha Gundapuneedi: to first discuss and then communicate with the team Yeshwanth Reddy Yerraguntla: No, no, that's a good idea. What I'll suggest is I'll add one additional u constraint there because everyone has their responsibilities. I want everyone to communicate in an async fashion. I keep using that word. Rajashekar G: Good Yeshwanth Reddy Yerraguntla: Sometimes I say you know just create this in viasa and give it to me. Rajashekar G: morning. Yeshwanth Reddy Yerraguntla: Right? Similarly whatever additional information you have created uh you have created uh have those links here. So we here we are saying are we ready with user stories uh simulation and user stories right whatever is that link I'm not saying I want vasa link I'm not saying I want only whimsical link whatever you think is convenient for you to just say hey I've done this these four things uh yesterday here is the link or here are the four links what we can do is Manisha Rashikar and I we can open up those links and consume them and understand what is what you're trying to do. 01:41:39 Yeshwanth Reddy Yerraguntla: That way our discussions can be uh including what you have done. Correct. Panel every day on that day. Give those links. Rajashekar G: You're right. Yeshwanth Reddy Yerraguntla: This everyone has access to. Oh my god. This has so many people. Okay. Why is so many people out Rajashekar G: This is in the like projects drive. Yeshwanth Reddy Yerraguntla: there? Rajashekar G: We can directly share the link. We don't don't need to add people. This will be have the access to everyone actually in our team also. Yeshwanth Reddy Yerraguntla: Oh, okay. Okay. Okay. Rajashekar G: This is in projects uh share drive. Yeshwanth Reddy Yerraguntla: Great. Yeah. Wait, what? done. If you guys are clear, you can drop off unless there is anything else I'm forgetting. Rajashekar G: And at 9:30 we're having call with Yeshwanth Reddy Yerraguntla: I know I know the we'll have to drop off I have to prepare I don't 01:43:02 Rajashekar G: VJ. Yeshwanth Reddy Yerraguntla: know he took 40 15 minutes long actually you know what let's drop off uh Manisha we have a meeting with uh some external person Rajar and Izza Rajashekar G: She she connected in the evening. Manisha Gundapuneedi: No, Rajashekar G: Uh uh she didn't join but she know like uh we had a discussion Manisha Gundapuneedi: I didn't join. But then uh Okay. Rajashekar G: while Manisha Gundapuneedi: Are we connecting back after that Yeshwanth Reddy Yerraguntla: We will connect back after that. Manisha Gundapuneedi: call? Yeshwanth Reddy Yerraguntla: Probably we'll connect back after uh whatever 1 hour gap is calling. Manisha Gundapuneedi: 11 hour 9:15 1 hour 45 minutes. Yeshwanth Reddy Yerraguntla: Yeah. Yeah. 9:30 uh 11 connect Manisha Gundapuneedi: Okay. Because we didn't give you updates on what happened yesterday with Navin and also Rajashekar G: Oh, Manisha Gundapuneedi: on a Yeshwanth Reddy Yerraguntla: I know I know yeah yeah yeah but we can't uh I can't avoid this Manisha Gundapuneedi: sure Yeshwanth Reddy Yerraguntla: this is at 9 9 and 10:30 anywhere between 10:30 and 11 will Rajashekar G: heat. 01:44:07 Yeshwanth Reddy Yerraguntla: connect sir Manisha Gundapuneedi: enough we'll reach the office. Rajashekar G: Uh guys, um Abilash this needs to be fixed as okay four points I'm sharing Abhilash Adunuri: Look. Rajashekar G: no already prompt Salesforce Abhilash Adunuri: Yes. Rajashekar G: prompt. I'm sharing my screen. already. See Okay. Salce any questions we ask like most of the time it should give distribution last month. It should automatically give the visualization and insurance flag. Hello. So Amulya Maggidi: Yes, Rajashekar G: actually Amulya Maggidi: sir. Rajashekar G: so we need to have insurance DB plus Salesforce agent in enterprisegami.com. Manisha Gundapuneedi: Shakra this can drop off right. Rajashekar G: All right, Manisha Gundapuneedi: Thank you guys. Rajashekar G: Harit. Arpit Pathak: Thank you everyone. Manisha Gundapuneedi: Thank you Rajashekar G: Hash flickering Yeshwanth Reddy Yerraguntla: Goodbye. Rajashekar G: issue graphs text. Okay. the ide. Okay. Huh? Amulya Maggidi: insurance merchant to Salesforce. Rajashekar G: Uh, Salesforce Amulya Maggidi: Okay. Salesforce. Rajashekar G: already. Amulya Maggidi: Um, and ch. Rajashekar G: Whatever the branch we have used for de enterprise brain Amulya Maggidi: Okay. Rajashekar G: perfect. Amulya Maggidi: Done. Okay. Rajashekar G: and um Abby what suggested is I'll share you one output actually flickering issue. Yeshwanth Reddy Yerraguntla: I'm Rajashekar G: So e Yeshwanth Reddy Yerraguntla: back. Manisha Gundapuneedi: Just explain them why we are see you are giving the points but Rajashekar G: JSON. Manisha Gundapuneedi: okay. Yeshwanth Reddy Yerraguntla: This month Rajashekar G: So Abhilash Adunuri: Oh no. Okay. Rajashekar G: okay and why this is because leaving to us today evening. So before that she want to see everything is right. questions. Okay. Abhilash Adunuri: understood. Rajashekar G: Thank you guys. Kalakonda Harshith Rao: Thanks. Thank Yeshwanth Reddy Yerraguntla: Yeah, Raj, we'll connect in a different call. Transcription ended after 01:50:40 This editable transcript was computer generated and might contain errors. People can also change the text after it was created.