कुछ बोलिए ना। Yes, यार, कर दो।

अरे भाई, हिंदी में भी yes, कोई-- हिंदी में भी समझ पा रहा है। अरे भाई! समझ रहा है, तो बहुत बढ़िया है। अ, let's now talk like the use cases, like पहले यही वाला like let's, let's just reiterate it. हाँ। So, essentially what, what we want to have, like you said, you say it. हाँ, so, basically if I am consider I am a husband and I have a wife. So, suppose I am supposed to go out tomorrow grocery shopping के लिए and my wife knows that I am भुलक्कड़ हूँ, भूल जाता हू�

�, तो वो इनपुट ले लेता है और फिर वो जब भी वो घर से बाहर निकलता

है। ये, ये, तुम बाहर जा रहे हो। Looks like you are going for grocery shopping. Here are some requirements from your wife. तो वो उसको बोलता है, "ये तुम्हारी वाइफ की याददाश्त है।" और ये बेसिकली तो पहला use case थी ये वाली। Essentially like, essentially having, having this to do list- हम्म। ...updated with user का routine का context- हाँ। ...and then reminding him with a notification that things- हम्म। ...need to be done. और फिर उसको याद दिलाता है कि चीजें इस तरह से जुड़ी हुई हैं। हाँ। और, और हमें ये देखना होगा कि, अ, हमारे पास permission control settings हैं। कि मतलब मेरी वाइफ तो मेरा to do update कर पा रही है। कोई random बंदा नहीं कर सकेगा। मतलब ऐसे- हाँ। तो, ये कैसे काम कर रहा ह�

�? अ, if you create two connections between two people- हम्म। ...it will have a, it will have a, um, unified, uh, place of this thing. Um, it will have basically a shared context. हाँ। Each shared context will have like-- shared context will exist for, uh, like two, two people. हम्म। One-- like, there will be two, two instances of this. One is like my shared context with you- हम्म। ...and your shared context with me. With-- हाँ, okay. Each of our shared context will have certain permissions that we will be able to exchange between us. For example, like-

हम्म। ...can this guy access my location? Can this guy set reminders for me? Can this guy do X, Y, Z? Like, we can have certain, uh, number of, like, things. And, um, based on this, uh, these things will be-- anything that is agent to agent is-- happens on the background via-- primarily, I am thinking via notifications. And then, like, certain, uh, rundown of those notifications in the app as well. Um, for example, let's say ki, like, like, you tell me ki, you want to meet me, uh, uh, let's say, like, on Fri-

Hmm ...day. And we have, like, this friend ka connection already. Toh, like, we can have, like, set up a meeting, like, sort of a thing- Hmm ...wherein, wherein, like, um, when it, like-- you basically come, send a, uh, meeting invite to me at XYZ place. My agent has access to my calendar, and my agent also has access to, like, talking to me at the end. Hmm. So, instead of you calling me and you thinking ki, what will be done, first layer, calendar will be sorted. Second layer, uh, I would be asked directly-

Mm-hmm ...also whenever I'm casually chatting to the agent. Ki, "Riteish was talking about, like, meeting you on, uh, in, in, uh, Airo City on Friday." Hmm. "Are you-- You seem to have, like, a lot of time on Friday. Maybe post office, you can jump there. I can arrange a meeting there." Haan, haan. Aisa, something like this. Yeah, yeah. Wherein my agent also knows daily when I go to work, when I come back- When, when you come ...what sort of activities I'm doing, uh, daily. We can also, like, talk to u- the user and basically ask him about his daily routine. What does he do? What does...

Yeah. And this will be done by a, a onboarding conversation, which will continue multiple times so that we can update our memory and- Uh-huh ...like, keep this. Uh, uh, uh, mere dimag mein ek aur use case aaya. Matlab, it is not personal rated, but it is mostly work related. Ki, uh, I'm seeing ki, like, aise Deepansh, he, he has to ask for updates from people individually. Aapke, uh, sabse matlab individually dekhe, updates se nahi padti. WhatsApp update, WhatsApp update, WhatsApp update. Uh, what if he just tells his agent subah-subah ki, "Dekho, Manav ko yeh karna hai, Ekash ko yeh karna hai aur usko yeh karna hai." And then I get-- I-- mere-- hum sabki mein alag-alag team bana diya.

Okay. Two-two add ho jaaye, theek hai? And then at 4 PM around, uh, hamara agent hamein ping karegi, "Tumne kitna kuch kar liya. Kar liya-- yeh, yeh ho gaya, yeh ho gaya." And then we just tell him ki, "Yeh, yeh ho gaya, yeh rehta hai," vagairah. And then that mo-- memory is updated on Deepansh ke side, so he just asks his agent ki, "What's the update from these people?" So, he just gets update from all the people. Haan, so we can have, like, a work, work connection also. Haan. Without, without actually having all the bullshit of, like, organization support, this, that, yeh, woh. Yes. Like, Deepansh aise five people are working under him. He just goes in the morning and he asks ki-

Mm-hmm ... "I want updates on this, this, this, this, this." Yes. And, uh, "This is-- was assigned to this guy and this was assigned to this guy." Yes. "Please go and ask and also ask others, uh, what they're working on." Yes. And then our agent comes and gives us a notification, like, "Hey, uh, what's your- Yeah ... what's your update and what's your status on the task?" Yes. "Maybe talk or, like, maybe..." Yeah. And that is then reflected to Deepansh. Yep. How do you think, um... Then, then how do you think actually the, the flow of conversation should work? Ek toh yeh ho gaya ki hum kya, kya karenge, haan, because-

Mm-hmm ... a lot of things can happen just by this simple shit wherein, um, like, I have basically unlocked this method of you say something and a notification is popped on my device. Mm-hmm. Right? This sort of thing I have already achieved. Toh- Okay ... what this can help us, let's say I go on, on the app, talk something, and based on certain, like, triggers in this conversation, I can pop a notification on your app. Yeah. Basically. Yeah. Because we are-- considering we are, like, already connections. Toh-- but now my, my issue with...

This sort of thing will be, okay, notification like, what d- what does it open? Is there a voice? Is there a, is there a text? Is there a something else? We can have both, but primarily we can f- uh, focus on voice. start recording or start your reply. Uh- Uh, I, I would generally want th-this, like, uh, like, as a scenario of, like, let- let's take an example scenario. You, Deepansh, like, let's say Deepansh, who's basically, like, taking updates from us, he asks me, "What is your update?" Mm-hmm. Uh, I-

Mm-hmm. A notification pops at certain time, like, let's say every day 1:00 PM or 11:00 PM- Yeah ... something like that. Um, I go click on that notification. The agent starts playing, uh, like, the audio conversation starts- Yeah, yeah ... right then and there. And I basically tell hi- tell the agent like, "Ha, ha, ha, this is my update, this is that, that, that." Yeah. And it instantly converts it into text and sends the notification back, considering the context of work. Uh-huh. So, this sort of, uh, thing, uh, uh, uh, I was thinking you . But most-

Yeah ... like, he is . No, no, we can have a scheduled notification. Yeah, yeah. Uh, what happened to, uh, uh, uh, task. Hmm. Yeah, instead of, like, sending five notifications to Deepansh, it is better that, uh- He can just ask. ... you just send one notification- Yeah ... to Deepansh on today's updates. Yeah. And then today's updates are, like, kind of-

Yeah. ... like, accumulated from all the work connections that we, he has. And the agent comes and plays a recording of that when he clicks on that notification. Yes. Um, recording after summary. Could you mean that? Sum- Yeah, yeah. Like, not... It's not like our voice. Yeah. Because, like, the agent tells him- Agent. ... "R- R- Rithik said this, that he'll do this by this point, and then this." And basically, like, the tone of the agent can change based on Deepansh, like, . Yeah. Yeah. Or full context . Yeah. That can change, like, if he asks, like, uh, ask more about this to Rithik. Mm-hmm.

And then, uh, like, later on, things, like, questionnaires starts. If we can somehow... Oh, . I think, yeah, agent to agent is something a bit unique. . Agent to user, . This, this, what we were talking about is agent to user. Yeah. Or user to user, basically. User to user via an agent. Via an agent. Yeah. . Instead of a single agent being there. I mean, essentially, it's not like two agents. It's like how I'll basically

tell you. Uh, based on ce- certain permission level accesses, , I, that I have on you. Mm-hmm. Right? Maybe it's not, like, calendar or something like. Maybe it's just, like, simple thing like location or something. I will have the per- ability, like, I'm giving the, me the ability to ask your location without actually asking you because you have already given me that. Given you, yeah. Right? Similarly, let's say you f- we'll have to kind of feed a lot of data inside our, uh, memory. Mm-hmm. So-

Mm-hmm. ... the next goal is I'll basically re-architect the entire thing, like create a entirely new project- Mm-hmm. ... which will, which will have, like, uh... I mean, I'm not really sure how this will be architected eventually, um, or how the DB schema will work, like, but we can have, like, something simple as for the POC- Yeah, yeah. ... where, you know, it's just a chat with, like, people talking and, and we, we don't even consider, like, we don't even consider, like, the possibility of connecting to calendar, . Mm-hmm. Let's just keep it voice-driven and action-driven from the user. Mm-hmm. Once we have done that, it will-

Mm-hmm. ... be very easy to plug in, like, things like, um, uh, Google Calendar or . Right? Yeah, yeah. For now, let's just do, okay, I ask you something and something is updated. So basically, it's, it's like, like the rem- like the grocery thing that you explained, right? Yeah. I want, like, um, I want some updates from someone, but not right now. Mm-hmm. I know, I, I remembered, I know right now, so yeah, maybe they can update in the-

Yeah ... later, later on part of the day. Yeah. Or maybe I want, like, the week summary of, like, whatever updates were given. Mm-hmm. Yeah. These things can be stored in memory and in context. Uh, but how much will it cost? Like, probably too much, like- ... you know? It's, like, going to be an expensive lot. Uh-huh. If you, if you just, like, go to Meme, Meme Zero, Mimo, I mean, whatever it's called. So...

And we are obviously going to use their API.

It's going to, it's going to cost us...

20,000 retrieval requests and, and 200,000, uh, ad requests, which is like ad requests will be less. Retrieval requests will be higher. Okay. So 20,000 re- I mean, one conversation will have like at least five. Five? Yeah. If the user talks to this like five times a day, so that's like...

Let's just say 50 per user per day. Mm-hmm. And, um, 50, and let's say we have like 1,000 users. Okay. Like, like a generous amount of 1,000 users. So that is 1,000 times 50 is five lakh. Yeah. And five lakh times- No, no, 50,000. Uh, 50,000. Yeah, 50,000 times, uh, 30 days. 10 lakh. 10 lakh. Yeah. Yeah. For, for like, for like 1,000 people-

Yeah. ... a monthly cost around $100 is like... And it's not that expensive to be fair, like- Uh-huh. ... for 100 users, like $100. Yeah, yeah. Yeah. Less than $1 per, per user essentially. That's pretty fine, I guess. Uh, so it's manageable. Yeah. Uh, we can also do this thingy, certain things we don't even put inside the memory. Mm-hmm. Like, like we know, okay, this, this guy does this every day. Personal information, though, we can

like, based on that pattern. But how will we analyze those patterns and how will you make a sub-learning? It's like a discussion for, like, a bigger- Uh, i- it's like temporarily stores in short-term memory, but over time, I mean, but this is like... Huh. Yeah. Like the meme API is like kind of- Short-term. Not even short. Like, it's like- Mid-term.

... short-term memory and then, then, then if we store it in our DB, it was like hardcoded, like, you know? Mm-hmm. Mm-hmm. So, like, we can have, like, just a simple text file, dump there, like, kind of do over, let's say, every month, once a month, we do, like, we, we do, like, five- We just- ... calls. We just, like, we just call the entire memory- Mm-hmm. ... out, out, and then we say, using an AI, Hmm. ... highlight-

Mm-hmm ... the important ones, and this is the user's per-- like, default, like, personality. Mm-hmm. And based on this, like, match and update the personality- Yeah, yeah ... bit by bit. Yeah, yeah. Yeah. You know? Yeah. Uh, and, and keep these things separate. Like, we can also have, like, a default personality that user gets, gives us during the onboarding. Mm-hmm. Where we can have certain standardized questions. What does he do? What does-- what do you eat? What do you... Like, kind of getting to know stage. Uh-huh. I, I guess, yeah, though, uh, but main thing is to think about the use case. The use case, we have to think about-

Mm-hmm ... the use case. And according to that, I will design in the same way. Because the problem now is that, uh, the design I showed, it works in minimal use cases, but if the use cases are big, then it kind of breaks down. So I'm not sure what will happen there. So- I mean, don't try to expand the use cases in the sense of, that we will do a lot of things. Rather than that- One, uh, I mean, one functionality can be used in different ways. Yes, yes, that is what I want. Yes. Like this, this, uh, this thing that we were talking about, reminders. Can this be used in multiple places like, um-

What if, like, वैसे instead of a notification, वो कोहली मिला दे तुम्हें, ऐसे agent तुम्हें call मिला रहा है कि अरे भाई, यार ।

Agent ने कोहली मिला दी भाई, याद है ना वो? कोहली कोहली कोहली।

हाँ। Calling agents can be set up also, like, yes. That's not hard actually. That's already figured out. text and like it will do it. उससे ये होगा कि क्या बोले people-- अगर उन्होंने notification ignore कर दी तो ऐसे call करके उतार देगा। But anyways, getting off the point. क्या कर दिया use cases होंगे? और एक और use case मतलब जैसे कि कोई ऐसा जो अपने आप को बताए कि अगर मेरे को ये काम करना है तो मैं कर सकता हूँ। तो उसके लिए एक ऐसा जो आपको बताए कि आ

प कर सकते हो। तो उसके लिए क्या कर सकते हो?

नहीं, वो-- how does the algorithm-- इसकी problem कहाँ है? अ, what, what, what I have a problem with this. हम्म। How does the architecture look like from-- on the backend? मतलब, because just like the agent is like one part of it and the memory is one part, then h-how is it orchestrated? That is another part. तो अभी पता नहीं, अ... What I am thinking is, see, every person gets a durable object का, अ...

Can-- let me search actually. Durable check.

Okay.

Yeah, I guess, like, what we can do is, let's say, I'm, I'm deployed on Workers, like, every time you come on the app- Mm-hmm ... you are allotted a separate small durable object instance, which, which Cloudflare does some magic on, and it kind of saves your storage or and it pops your instance, like, every time. Yeah, yeah, that's like, bro, at the end, this is exactly what I need to do.

And I feel bad because, like, it's, I don't think it's a good enough product. Yeah, yeah, yeah. It, it- See, whatever we are talking about, at the end, just a call from my wife works. Uh-huh. Basically. That's the fucking problem. Even, fuck it, a WhatsApp text and a call from my wife works. Yeah. That's so much lower effort.

Mm-hmm. And it doesn't cost, like, a penny.

This shit hard to figure out. It's a use case. I'm not... Nah. See, that is the-- that is not our problem. That is a problem in the market itself. Uh-huh. Yeah. There is no...

There is no use case for this sort of thing, wherein...

I mean, sure, if you guys want to build it, like, anyone can build, like, a prototype and some booga-booga around it. But like, yeah, I mean, whatever. But, yeah, AI, it's easy to get funding and then get users, at least, uh, users if you throw at marketing.

हाँ। नीचे पे 500-600 बंदे आ जाएंगे, आराम से, प्यार से। Yeah, yeah, yeah, basically. ठीक है? पहले दो तरीके के लोग हैं, तेरे normal लोगों को भी है ये-ये hype है कि हम भी अपने tooling में कहीं डाल दे। हम्म। ऐसे पे पापा हैं।

Voice note 2026-05-08 13:51.weba