Cognitive Cell: Building a Route-Select-Render Control Layer for Workflow AI

Speaker 0: Sai, yeah. Speaker 1: Yeah. So getting Sai, this is like a project that I AI Mike recently started. It Cell like probably like not even like a v Jun. It is like v 0 point something. So, the idea is how are we meeting the AI autonomous in the sense that it has like more Control, right. Speaker 1: Sai for example, we are having like Robb Coding, but the thing that I have observed is why is City? So, why are we like even prompting it to get started AI. Sai, if I am wearing a hat of a data analyst or data scientist for example, I would probably Pun the laptop, probably look at some Jim tickets or some dashboards Render went down, I would probably investigate it AI. So, I mean that is the idea that I have using to you know, using as a lead to developers this 1. Sai, what is does systematically, City, right now, does this very simple task of using something. Speaker 1: So lead us say using AZTRA something. So it basically retrieves the context AI, say for example, a blue color is observed. So what is that means to do? So it means difference in different contention AI? Say for example, it can like observe the record and do nothing or probably escalate it or you know, do something else. Speaker 1: So, Cell three actions have been AI identified AI like categorized into AI, Robb, I would say, 6 exhaustive categories for now, which is like record, classify, analyze, balancing, direct search, and escalate. So when I say this categories, this is currently for like a workflow AI and probably a corporate meeting, but it can also be expanded to further Vue know sections or further industries. Sai that is like the 1st part which is like the router Jun the 2nd part is selector. So selector is CLI, what, it needs to do. AI choose the best pathway which is like workflow style response or direct search response. Speaker 1: Direct answer response is like you know, hitting an API Tools answer City. And now, finalizer is where things get a little bit interesting where ah. So, what I have observed is, whenever you have something the traces and the workflow records are useful internally, but they need to be Render clearly for the human to use it. So that was the idea that Live used to basically, build this thing, like a Pun Cell, but the north star, is basically, to replace this part AI the user has something has to go off. So, for that what I am trying to do is build build a Robb a self model that kind of understands, what is the JSON or what is the contention kind of adapts to it. Speaker 1: So, AI mean Route now, I have just published this package. City is called Cognitive Cell. It is on AI Jun you can probably try it. So, the context over here is something like this. Say for example, there is like a warehouse Event. Speaker 1: AI hope the font is fine. Say the package label shows the wrong city. Sai for example, that is what is the statement or the user prompt. So, how could it Real? Like we demonstrates Actions goals AI etcetera Center Jun these together probably would, Vue know, Date for three next response or the next best thing to do. Speaker 1: The same thing in a medication event probably has to be responded in a different scenario. AI? Sai, the idea is to make it so autonomous that Vue need not even prompt multiple times that you shouldn't do this or there is some caveat. Probably if you are working in a Health care setup, there might be a lot of AI and stuff. Right? Speaker 1: So the idea is to basically have all those predefined or even Live, Searching pre Engineer, so that it could just act on the go. Sai, that was the main idea Jun yeah. So, AI mean as I have said, I have done some validations or like how it is better than Mike a plain charge GPU Sponsors. Jun that AI think out of Mike 100 cases, I got 2026 AI. It's better than GPU Jun. Speaker 1: GPU 5.5 Cell Live 55%, but as I've said, this is AI 2026 even like v 1. JSON, and I've done Live a Jun event enterprise sidecar pilot. So, I basically evaluated few metrics, and it said the useful 1st move is based on this particular architecture. And AI, I have also performed a few other things where, it is AI Date to how can this be compared to the AI, which is baseline AI like GPT-5.5 or GPT-four Android of 100 cases, 79 cases was basically preferred this particular thing to be AI more contextual or more meaningful than AI a plain chat City Sponsors something like that. So, yeah, I showing, this is the progress for now, but showing forward, what I'm planning to do is scrape some of the data, try it Street out, like, with more use cases run, give autonomous control of the systematically, we can evaluate the next steps. Speaker 1: So, I mean, in the process, there was obviously some things that went south. So 1st thing is, so we the thing that I came up with, like, the 3 Street, which is AI route Select 10 finalizer. Basically, Actually, the route got over AI, which means the earlier versions were just Mike asking a lot of clarifying questions, which kind of increase the API cost as well. So, AI thought I thought like showing it downward Mike sense. So, and workflow architects were also like too raw. Speaker 1: So, probably having like a pretty strict JSON Sponsors, Live me Real, Sponsors. Search, yeah, I think the AI, the chat Sponsors were also stronger than expected. But, once I came up with this particular, architecture, it kind of solved things initially. I mean, at least Live me a direction to go. So yeah. Speaker 1: I mean, probably the next time Sai Mike present, this might not be even Silent, but, yeah, just, telling how I'm, trying to build these things. Speaker 2: Yeah. So I had a question on, the premise was AI user questions, but it seemed like your example data was like an event-json? Speaker 1: Correct. So, so the example data was like, there's, like, an observation. Speaker 2: Okay. That's like separating Control. Speaker 1: Yeah. Supporting context or something Winkle that. Yeah. What's the next best thing to do? Like, even if we Sai AI City, Mike, it's raining AI. Speaker 1: So it could give, like, meeting. Like, wear an umbrella or even AI, make sure something is outside. So, the idea is to make it too weak and the system has to respond in a way that it already kind of, takes the context not AI, you know, telling a lot of AI USA of prompt-generated Speaker 2: it is routing, what is City deciding between to route? Speaker 1: Sai, CLI City appreciates the situation based on these 6 categories for now. So, record, classify, analyze, Layer, Silent answer, escalate. So, based on this City kind of triggers Vue showing, which is AI, AI mean I am currently Vue know the demo might not Workflow now, but the idea is to basically trigger like a JSON Event of doing the next best thing. Yeah. Speaker 2: So it's in it's intent. Speaker 1: It's intent. Yeah. It's intent. Right. Right. Speaker 1: But how you I mean, if if it if it were given access to the systems, you know, I could probably, evaluate it a little LLC Center. But, for now, I thought Live, ah this is 1 better way of going yeah. Speaker 0: And what do you Route it to an SLM or LLM or? Speaker 1: Yeah. So, I have tried it with a lot of large language models I have tried Pun AI, AI have the LLM, I have tried ah ah CLI and also AI. So, mostly most of the responses were kind of similar Sai have also done some evaluates based on for each of the lessons how the responses were. So, GPT responses Cell kind of consistent over you know few of these observations. So I thought, like, Jupyter is, like, good for now. Speaker 1: Yeah. Yeah. Speaker 2: Are you, when you're doing this, do you do you record it through a log at all? Because I feel like this could be useful for possible, like, routing and fine tuning and things like that. Speaker 1: Correct. Speaker 2: I I issue I have a lot of issues when it comes to steering, you know, when you get to a big Speaker 0: thing that you're working for. Speaker 1: So Yeah. So you're asking, like, am I logging it off, like, in memory? Speaker 2: Yeah. Like, when you so when you're recording Meetup, like, the the routes here, are you are you, like, is there are you kind of attaching it to, like, you know, keeping a record of everything that you AI? So if you wanted to, you could feed that Yeah. In this Speaker 0: training either. Yeah. Speaker 1: That's that's the whole idea secure, over time, it has to become so robust that GitHub even telling anything, it has to, like, act upon it. Speaker 0: Sai, Speaker 1: yeah, so currently, it's kind of stateless, but what I'm trying to do is adapt memory as well so that it becomes more more nuanced. You had a question? Speaker 0: Yeah. So the question was it it sounds like you're trying to this is what I'm trying to understand. Are you trying to create a a way that, whether it was a it's a worker or AI don't know how you plan to implement it. Speaker 1: Yeah. Speaker 0: But it's you're trying to get whatever, the Presenter is how somebody would respond to it if it was going to ask you clarifying questions Correct. Or anything. And then you kinda like that digital 20 yourself or Right. The profile. Speaker 1: Right. Right. That's that's exactly the persona is kind of yeah. Speaker 0: And then the question then would be, when you say clarify for the router, is that, like, just is that similar to a Sponsors, but it was more a selection? Or are you thinking model, like, it's gonna ask talk? Because I think Vue you just want a single output. Speaker 1: Yeah. That's a that's a good point. So the clarifier is basically for human intervention for sure Speaker 0: because, Speaker 1: or if you basically, look at this, this is basically like a lead block Vue could imagine. Sai, probably author lego block would clarify it. So, so, but, say a lego block a should ask a lego block b that it needs clarification. Sai City could be that or it could also be a human intervention if it's probably pretty critical or, you know, yeah, important. So yeah. Speaker 1: Yeah. Speaker 0: Yeah. Yeah. Speaker 1: I mean, there's, like, completely architecture based approach for run, but, you know, you can add add, like, everything to City. Like, Sai, there's CLI the way I developers it is looking at how can we make it so atomic. So lead me actually showing, yeah. So, I Date this as lead, so that, all these things are like pretty atomic. So, you can add like whatever API calls Vue Live, whatever routes you have. Speaker 1: So, real time context also yes. Yeah. You can do that for sure. Speaker 3: It can be either the same. And keep asking questions. Speaker 1: Yeah. Yeah. Sure. Speaker 0: You can keep answering we can keep taking questions, but I'm just gonna have to tailor. Sure. Absolutely. Yeah. Thank you. Speaker 0: Yeah. Speaker 2: Any more questions? Speaker 1: Yeah. Speaker 2: Just the router using structured prompts in Speaker 1: order to determine which type of not exactly structured Pun. There's AI a a Markovian based Route based methods. Basically, that's what is being used AI now. Yeah. Speaker 0: Thank you. Thank you.