Full Transcript

Utsav Somani: of AI, like the specific tools that you can use, build actual workflows in your daily life and I think improve your daily productivity by a mile. All of this we've of course read on Twitter, we've seen on LinkedIn and people making trading bots and different things but how do you actually go about building this and the various tools that are required and also just improve everything in your daily work life. So we will welcome Vishesh to the show to run us through all of this as our first AI masterclass. Vishesh, welcome to the show.

Vishesh Khurana (Shiprocket): Great, thank you.

Utsav Somani: What's keeping you busy these days? Shiprocket is preparing for an IPO, we read Tribe Capital India, we read Kraken, so how do you find time to do all of these things and then also build these awesome AI things?

Vishesh Khurana (Shiprocket): Yes, for me I obviously took a small like operational back step with Shiprocket. This was the first time I wanted to be ahead of the curve in terms of 15 years of being in the startup ecosystem. Every time I think I missed the bus in terms of internet happened, e-commerce happened, we were little like late marketplaces, D2C. So this time when AI kind of popped up, I wanted to be very very ahead of this one at least. It's a moment once in a lifetime thing kind of happens to every entrepreneur. So I obviously took a back step from Shiprocket, spent some time traveling, moved out of India, went to Dubai, spent most of my time in the US where most of the action was happening and just trying to understand what people are doing and what people are building. Met a couple of incredible founders who were able to build half a billion, a billion dollar company with no employees, with no external funding. There's one guy who built a medtech startup in SF, sold it over a billion dollars with no external funding, zero employees, single guy, now valued over a billion dollars. That's the thing, so obviously had to implement a lot on my workflows with various things running. So I have been way more productive with adding these AI workflows, automating a lot of my daily tasks, my emailing, my research, finding, networking, whatever I do, having these automations in place now have made me super productive in terms of how I perform every day.

Utsav Somani: How much is your Claude Muncy bill?

Vishesh Khurana (Shiprocket): Oh, it's gone but this month was about 3000. So I've actually now found a hack to reduce that. It's three to four thousand dollars a month now but I've kind of found a local model. So I've just installed local models on my laptop trying to now figure out how to reduce it. It's a small cost but still we are always going to be cheap. So I'm trying to reduce that also.

Utsav Somani: What are you using currently? Like I mean, what are the 10 tools that you've used in the past one week?

Vishesh Khurana (Shiprocket): My most, I'll go back on daily. So every day I will spend some time for research. Let's say when we have to do deep research, of course there is deep research functionality within GPT, within perplexity and Claude and all. Somehow I have found the best performance by Kimmy. It's a tool that's, I think it's a Chinese tool called Kimmy. So if you go on agent swarm, Kimmy has a, let me just share my screen. So while we talk we can also.

Utsav Somani: We can, I mean share your screen and just keep showing. Like I think the idea is for us to learn.

Vishesh Khurana (Shiprocket): While we talk we can see. So that would be helpful. So this is a tool Kimmy.com. I think it's from China. And if you go down on this basically called what model to use, not for everything agent swarm but this is a very incredible functionality.

Utsav Somani: I mean for a non-technical person like agent, what will you describe it as? Like I mean in one simple line basically.

Vishesh Khurana (Shiprocket): Agent is nothing but assigning a task to one person, end of the day it's one brain. A swarm which as the name suggests is multiple brains working in parallel. So if I have to give multi-step tasks, for example a task that requires first research, then find, then make, then code. So if it's a multi-step process. So let's say I have a task of building a micro SaaS app. So first I can go to Kimmy and say here find me. Basically I can just tell Kimmy to find me a best performing app on Shopify. Okay. Then tell me the revenue, tell me how it works. Then also tell Kimmy to code but I will not use Kimmy to code but it can do it. That's the point.

Utsav Somani: Basically it's running like non-sequential tasks in parallel.

Vishesh Khurana (Shiprocket): Yeah. So non-sequential tasks in parallel and it can spin out at least 100 agents at the same time. So you can give it a task and hundreds of agents will work together to give you the best outcome and it comes up with like, let's say this was one of my tasks. So we can run one in parallel but this is what I have done one of my tasks. So idea was that I have Shiprocket. Tell me what apps can I sell to my current base of customers on Shiprocket by researching on the best performing apps on Shopify. So now you see the table there. So what all tasks cover basically? It's basically analysis of what Shiprocket is, what market opportunity is. It's given me 1000 micro app ideas like a thousand apps that I can build on Shiprocket. On top of that it's given me 50 priority apps that these are the first 50 that I should be building. What these apps can build as a revenue? What is the financial protection? What these apps on Shopify make as a revenue? So coming back on what the outcome was of this task.

Utsav Somani: It can actually build these apps for you also, right?

Vishesh Khurana (Shiprocket): It can build these apps. So this was my outcome. So let me just open this as a sheet. So if you can see this presentation that it compiled for me, it's all done like 10-15 minute tasks. So it's basically, let me open it in the full screen.

Utsav Somani: But also you've been using Perplexity computer, right?

Vishesh Khurana (Shiprocket): That's, I'm coming back to that next. That is the most powerful tool, Perplexity computer with its browser called.

Utsav Somani: Perplexity just crossed 500 million ARR because of that. I think it's finally.

Vishesh Khurana (Shiprocket): The browser is incredible.

Utsav Somani: The search I think was getting commoditized. So now they've switched to this agents.

Vishesh Khurana (Shiprocket): Yeah. So that's been one of those. Yeah. Quickly, so I'll come back. So basically look at this presentation. It's gone back to, how do I close this?

Utsav Somani: I think we'll have to download the file. But okay, I think point it can do good research and stuff in parallel. You can have 50 of these running 100.

Vishesh Khurana (Shiprocket): Look at the amount of files that are generated. It's basically about 20 files. One is the compile, then each and every app, what this app will do, how this thing, I wanted to send an email to Sahil on how to, what he, we should assess someone. So it's created all of these files for me and I can integrate my emails and Gmail, everything else.

Utsav Somani: It can actually send ChatGP and Claude, I think it's single agent mode, basically.

Vishesh Khurana (Shiprocket): Parallel tasking on the background, they also run sub-agents, but for some reason I found Kimmy to be the most powerful of all of them I have used from a quality output point of view.

Utsav Somani: How much do you pay for it on a monthly basis?

Vishesh Khurana (Shiprocket): Kimmy is again about $200 a month, Claude is 200, Perplexity is 200. Each of these apps, the max plan is about $200, but I end up anyways going through my credits in a few days. So I have to add more credits almost every day.

Utsav Somani: You're doing a lot of byte coding as well. So let's switch to Perplexity computer. I think that also is super interesting.

Vishesh Khurana (Shiprocket): Yeah. So this was a Perplexity computer and the most interesting thing about Perplexity computer is the browser called Comet. It's actually, when you combine these two, it can do so many tasks within your browser. I'll give you an example. I built a small app for networking called Clia. And for that app, I wanted to now integrate Razorpay payment gateway. Now for getting the multi-step process here is, first I have to apply login to Razorpay, create an account, upload my company documents, get my account KYC, then take those keys from Razorpay, integrate with my app in Clia, test the payment flow, is the payment flow working or not. So I created a task sheet overnight. This was 11pm in the night. I just gave it to Perplexity computer saying, all of this, do it, use my email ID called vk.clia.ai to set up a Razorpay account. The difference between Comet versus Claude, I found was Claude stops when it comes to sensitive information. He's like, I'm not going to touch it. I'm going to not see passwords or anything else. Somehow Comet doesn't, even for captures, Comet doesn't care. It's like, they will do everything. So even capture cracking, it goes through capture like a human, everything. By the time I woke up, I had my Razorpay account set up. So I had to give obviously my GST documents and everything free. I uploaded to Perplexity on a Google drive before. It automatically went to my drive, picked up my documents, uploaded on Razorpay, picked up the keys, added back to my app. By morning, I have a payment gateway line. I can do this myself manually also, but why should I do it?

Utsav Somani: So run through this thing, what you're doing with Perplexity computer. I think that is interesting, but also Clia.ai. I think, what is the tool that you built? What was the prompting mechanism behind that? And this is like Bodhi, right? Where you can do intros.

Vishesh Khurana (Shiprocket): This started when I met Jivraj during one of our, just catching up and Jivraj said, have you seen this tool? I want to build something similar. Sure, let's build it. So we were sitting at Quora, let's build it. Let's try. So I just did a very basic prompt for this one. So I went to first in Chrome, Claude in Chrome. So there's that extension of Claude in Chrome, open the website Bodhi. The single prompt was reverse engineer this. That's it. One line. So it went through the website, understood what it does, how it was built, various features and functionalities of that platform, and everything compiled into a document. I took that document, took it to ChatGPT, asked ChatGPT to give me a PID. If I wanted to build a product, which is similar to Bodhi, but better, and add more functionalities, and also research what other functionalities would people of this app want. And then GPT went on the job saying, let's research on how to make it better. So it made a full blown PID for me. I took that PID. Now I went to Claude co-work, created a folder because now I have to integrate multiple GitHub hosting platform, everything else. So went to Claude co-work, gave that PID to Claude co-work to build this app in front of me. And that was the flow of how we went. I think by it took us less than 10-12 hours to come up with the whole app.

Utsav Somani: And it's fully functional.

Vishesh Khurana (Shiprocket): Fully functional, hosted everything. This is the website. Even the website, look at the design. It's okay. It's a decent design, right? Overall, like some animation, some 3D elements. Old times, these websites alone, a great job done kind of thing. So all these animations that show up. But this was the product. So all these LinkedIn integration, Google OAuth.

Utsav Somani: And have you installed any specific skills in Claude? And how can people go and find specific Claude skills? So let's talk about Claude skills also.

Vishesh Khurana (Shiprocket): Yes. Actually, there's interesting Claude alone was not as capable. But if you are on Instagram these days, you see these hundreds of Instagram influencers, AI influencers, who are every day showing some or the other tool or a GitHub repo or something or the other. So there are people now, as a community, people are making great tools on GitHub and making them available to public. Even these agents on my desktop was one of these, if you see this small two characters, if you see on my laptop, this was some guy on GitHub, just made it public. It's a GitHub repo, I took it from them. Now this is handy agent for me. I can ask, build a personal, let's say, personal finance manager, Excel template.

Utsav Somani: You can just do it from the desktop. And it's connected to your, you have to sign into all the accounts that you have.

Vishesh Khurana (Shiprocket): Everything is connected. It's all APIs and everything else. It's all connected. I can choose my models here. If I want to use Claude, Codex, Copilot, Gemini, OpenClaude, whatever. But single, this guy is giving me a very clean shortcut. So while I'm working, I can just give it a task. It is a GitHub repo, not too good. Personal finance manager, Excel template on my Excel on my laptop.

Utsav Somani: And have you installed?

Vishesh Khurana (Shiprocket): I've given it a task. You see this guy, it keeps me updated what you're doing. I can go back to my work.

Vishesh Khurana (Shiprocket): When it's done, it's going to open an Excel on my laptop and give it to me. Here is your file.

Utsav Somani: Look at that animation is working. Yeah. That is nice. But have you installed OpenClaude or any of the agents? I mean, people went crazy buying Mac minis and installing their agents, having them run 24-7.

Vishesh Khurana (Shiprocket): I was one of those people. I also went crazy. I have right now two Mac minis running with me. But you didn't need Mac mini. It was just an overhead. Yes, you needed Mac minis when your information is too sensitive and you wanted to have a closed loop information like a guarded environment system. I personally don't have any such information. I don't need Mac minis. I could just go install it into a virtual hosted server. And now there's hundreds of them. Even Kimi offers you something called Kimi Claude. It's a version of OpenClaude. I could just use this, a hosted version.

Utsav Somani: So you don't need to host your own basically.

Vishesh Khurana (Shiprocket): Mac mini is great because if you have sensitive company information, then you're actually going deep. Then you need to have a system that is always live. But now with Claude just today, I think or yesterday Claude launched a version of their own OpenClaude. So this thing is relevant now. Now Claude has their own version. They launched it yesterday called Routines actually, Claude Routines. So I can do the same thing. Give it a list of tasks, close my laptop and go.

Utsav Somani: And even that dispatch feature that Claude has is very nice, right? You're on the go, sometimes your files are on the laptop. So you want to co-work and access like, I don't know, a binary or like some PAN number of a company. You can just text it while you're on the go and it'll retrieve it on from your laptop and just give it to you on the fly. So I think that's fascinating. But I mean, all of this is also varying to the fact, I was reading about a case of OpenAI where somebody, one CEO asked OpenAI for advice and they actually implemented it, which was sort of unethical. And sometimes I think it was borderline illegal also. And they have the OpenAI chat history in the code document. So they've actually said that this guy used, CEO used chat CPU to get this advice and actually shut down an account. And they've pulled out the OpenAI history. So it's not privileged, like I think OpenAI and Claude. So that's why running your own bot might be something that you would look at doing, because then at least it stays with you.

Vishesh Khurana (Shiprocket): And of course, I understand, even when crypto, early days of crypto came around, no one understood what it is, how it works, no one regulated it. And over time, when we finally understood the power of it, and we have now regulators, rules and regulations and full blown ecosystem around regulating that whole thing. Same thing with AI, I think right now, no one understands how far, even today, because it's so easy, non-technical people are going on Instagram, seeing the steps on their terminal. There's some guy on Instagram saying, install this, then write this in your terminal, write that in your terminal, and press enter. And people are doing it without realizing because non people, I am non-technical, I also did the same thing. If you don't understand fully, it is scary because your data is now everywhere. So it will become harder and harder to protect your data. That is why, when I moved to Dubai, I joined this company, HALTASafe as an advisor on the board. My idea was, I wanted to learn from them more than they wanted from me. But they were building closed loop AI infrastructures for companies where it was hosted in their environment. So that will become a very big use case as we mature up in this ecosystem. There will be one or two big breaches and some big incidents that will happen before everyone just spins into action and starts regulating and creating guardrails. Claude just launched Mythos, if people follow news, vulnerabilities in the most systems on the world.

Utsav Somani: Decade-old operating systems and even broke out of a sandbox. So let's run through all the list of tools which are prominent today and tell me what are they good for?

Vishesh Khurana (Shiprocket): OpenAI? Coming back, OpenAI in general tasks, OpenAI is very good for reasoning logic tasks. So if I want to have some algorithms, like for CLIA, I wanted a matchmaking algorithm written for me. So I think I found OpenAI's codecs to be the right ones. It gave me the whole formula on how to, first of all, what CLIA is. Coming back, CLIA is basically a networking platform. On LinkedIn, we get lost. We all get thousands of spam, sales, all kinds of messages. The whole idea of LinkedIn was to expand your network and meet relevant people. And I also miss now, I get so many messages that I, important ones get lost in the mix. So CLIA is basically ideas to AI can now basically matchmake you, basis who you are, who I am, and if both of us are relevant for each other to actually connect. So that's a matchmaking platform. Now this algorithm of matchmaking becomes very critical. And this algorithm I had to basically open your right, all the tools coming back listing the tools, like Kimmy, great for day-to-day research, deep research. If you wanted to build presentations, research, the quality output is great. For me personally, I found the best so far.

Utsav Somani: But using it for research, like document generation and idea analysis and stuff.

Vishesh Khurana (Shiprocket): Yeah. So Claude co-work in perplexity computer for like a task. I just did this for a demo point of view. So I have a folder on my Google Drive where all of my employer, like people, let's imagine a scenario where all of your team members, finance, legal, marketing, everyone is sending you multiple reports and documents, what is happening in the company. Now you have multiple PDFs, Excel and presentation in one folder. I said, let's see, I put all of these documents into one single folder and say, I have a board meeting coming up. Help me prepare for that. Also, I have a meeting with SoftBank, help me prepare for that. And let's do some research on everything else I might need for these two meetings. Now it has multi-steps where Claude co-work and both can do. I found perplexity a bit better in terms of output of presentation design. Now this one will go through all the documents it has on my folder, take out all the information, compile it. Now once it's compiled all this information, it's spinning out documents. It's also done some research on SoftBank, who I'm meeting, where they invest, what kind of investments they do. Then it's now created a board meeting.

Utsav Somani: And perplexity is good for this research, but if I want to build an app, like I want to one-shot an app, like suppose I'm building the community app for offline. What tool would you use? Would you use Replit? Would you use Lovable? Would you use Cloud Code?

Vishesh Khurana (Shiprocket): I started with Replit. So most of my apps were built on Replit. It's the most easiest and the most powerful one. It has all integrations pre-built. The only limitation which is not going to be relevant for a lot of people is, as the app you have built on Replit starts scaling, let's say to thousands and thousands of users, the scalability of the backend infrastructure will collapse. But that's a problem for later. You can always move it out. But Replit is great in terms of...

Utsav Somani: It's good for white coding, but why do you need actual human intervention for debugging?

Vishesh Khurana (Shiprocket): So that's where you have Replit built. So let me take you to the stack. Let Replit build the MVP stage. Then have GitHub as a central system of hosting that code. So integrate Replit with GitHub. Now your code is on GitHub. Now, because it's on GitHub, I can now take it out on any other app. So let's say I can go to Cursor. Cursor is great for debugging and the quality of code is much better for me.

Utsav Somani: Talking to an audience which is non-technical. I'm just assuming that because some people will not be... I mean, the idea is for them to get comfortable with using all of this while being nontechnical.

Vishesh Khurana (Shiprocket): I'm also non-technical. So simple. Cursor is an aggregator of all models into one single coding app. So it has access to all models where it selects the right model for the right task. It can debug and improve the quality of my code and everything else. So I have my GitHub as a source, one single place where all my code lives. I can have some code written by Replit. If Replit can hallucinate sometimes, it fails to debug something. So you can go on Cursor, analyze my code. This is the bug I'm facing. Help me fix it.

Utsav Somani: So let's do this. Let's do a live demo where I give you an idea.

Vishesh Khurana (Shiprocket): Yeah, let's do that.

Utsav Somani: Build a community app for offline. It should have an events calendar, a members directory and a forum board. Okay. These three things. So show me how you make your prompt better. You type in...

Vishesh Khurana (Shiprocket): Hello, let's do a prompting, please. So you say, let me just give it a... I'll stop typing now.

Utsav Somani: Mostly it's whisper flow, dictation.

Vishesh Khurana (Shiprocket): Dictation. Build me a community app for offline. Understand what offline is on this link.

Vishesh Khurana (Shiprocket): www.offline.club. The app should have features like events calendar, a one-to-one chat for members. And what else?

Utsav Somani: And forum board, I think...

Vishesh Khurana (Shiprocket): Give me some ideas. What do you want me to do?

Utsav Somani: Forum board, this thing and personal profile of members, members directory that covers it. So I think this is good. Like, I think, let's see.

Utsav Somani: So now we're giving you the prompt, basically.

Vishesh Khurana (Shiprocket): I'm just lazy.

Utsav Somani: I mean, two years ago when ChatGPT came out, I think everyone was like prompt engineering will be a big thing. But you can actually, I think, use ChatGPT itself to tell you a good prompt for whatever you decide to use for building this.

Vishesh Khurana (Shiprocket): This is what now, look at the amount of context that GPT is given by one random line.

Utsav Somani: Because it's impossible for somebody to imagine all of this also.

Vishesh Khurana (Shiprocket): It just makes it faster. So it's giving you how it should be built, what exactly, what stack should you use, front end, back end. You just picked up everything. Now I can just take the whole thing and go to Replicate. Let's use Replicate for this one.

Vishesh Khurana (Shiprocket): Yeah. It's easier for me. I just go to Replicate. I pasted my prompt here. Let's go. So now it has a lot more context. So it won't ask me a lot of questions. It has a lot more context. So we can just continue while this does its work. I think 5-10 minutes later, we should have the first MVP of this running. Yeah. That's by way of at least how I would use this. And then I can actually use Perplexity Computer or Kimi agent forms to test this app. So the very interesting thing is once you've built the app, testing in QA becomes a very important one. Right. So if you look at this, I can show you something here. I can go to agent forms. I can have, so I have a founder console. I have a prompt written somewhere.

Utsav Somani: But okay. So I mean, tell me about the prompts that you built also. You built founderconsole.ai, clear.ai. But I think the most interesting one that you showed me when we met in person was your personal Slack summary that you get on Slack daily. Summarizes your WhatsApp, email, action icons.

Vishesh Khurana (Shiprocket): I have about seven different email IDs, given I'm involved in some form of capacity as advisor or something with some of the companies. Then I have my own multiple apps that I have built. So now I have like six, seven email IDs, almost a lot of four WhatsApp numbers. Here was how do I consolidate all of my information and make sure I am on top of my job.

Utsav Somani: But something, WhatsApp doesn't have their own MCP servers, right?

Vishesh Khurana (Shiprocket): Yeah, but you can still integrate using WhatsApp web API. So you can obviously pass through all of your WhatsApp communication through a channel. I built that tool as well. It's not live because I'm very worried about the privacy of this. But this is a tool I have built for myself, my own WhatsApp. It's not for public or anyone. So I'm still trying to figure out how to garden this. But idea was to, all my WhatsApp communications is passed through this. It reads all my communications. If you and me have decided, let's say, let's meet on Wednesday at 11. I have said yes on WhatsApp. This is just a one to one chat between you and me. It goes through that, it automatically reads that, sends up calendar invite to you and me both. Very simple automation. But yeah, it's a fun thing. But it has all the documents, everything analyzed. I hate WhatsApp search. If I want to search a document within WhatsApp, I'm always impossible.

Utsav Somani: Yeah. I think even in Slack, like I think in many other things like Notion and AI search is going to become default. Like Claude, you can connect with Gmail, Calendar, Google Drive, Google Sheets, Google Docs. I mean, it's impossible to find a document. Like sometimes I ask a coworker to just hundreds of files in a folder, find me that specific thing within that PDF. And it just looks for it and give me the answer in five minutes. So AI search will be big, even on your personal intelligence level. I think Geminal also.

Vishesh Khurana (Shiprocket): Exactly, on top of all this to understand that everything is organized properly. So I can go on my WhatsApp bot, which I built for myself, saying can you send me my mother's blood work from last year. So it will have access, find it and send it to me.

Utsav Somani: Show us that Slack bot that you built and what all it provides.

Vishesh Khurana (Shiprocket): Slack bot is basically, it's going through reading all my emails. So every morning, so I have to be updated. What's happening? It reads all my emails. What emails and summarizes everything and gives me a listing. These are emails you should, I should be responding to first. And also creates draft replies for all my emails. So all these replies are in the draft. So if I'm in a car on my phone, I can just open my emails, see my drafts and just send from there. So I don't write or if I want to make changes to the draft, I can do that. It actually creates those drafts for me as well. And over the time, it's now understanding me, it's also learning. It understands how I write, it understands what I do, how I reply to emails. So it's getting better every time.

Utsav Somani: Alrighty. So we had a minor technical glitch. Things that happened when you're doing this live. So let's jump back into the conversation. So we were doing, making an app. So I think, let's switch back to that because we want to build a live, this thing and you were showing us your Slack bot. So there's a question in our chat where somebody is asking, how can they build it for themselves? So can you tell us the steps and the processes to building a summary bot for yourself in Slack, WhatsApp, Telegram?

Vishesh Khurana (Shiprocket): It gives me access to only connect one email ID using the Gmail connector. So I built a dummy email ID, forwarded all my other emails to that email ID and now Claude can read all my emails. Then I have in the within Claude, you have Slack connector. So it has an MCB connector. So I can now have it summarized and send it to me on Slack. I've went a little bit one step further, where I also can reply back to that same Slack for agents to do any task. For that, I have an independent agent on top of it.

Utsav Somani: Is it a two-way agent or just a read-only?

Vishesh Khurana (Shiprocket): Yeah, so it's a two-way communication. So I get my updates like these with all my emails, deals, whatever it is, it basically takes down the most important ones based on my guidance. For me, I care for most time-sensitive emails first. For example, if there is a time deal that I should see right now, if there is a payment or credit card payment or something that I missed or some bank update or something like that. So for around that, it summarizes and prioritizes my replies and everything else on that as well.

Utsav Somani: Interesting. So I mean, people should build this for themselves. I think it's solid and you can choose different news items that you can summarize like sources.

Vishesh Khurana (Shiprocket): I'm also doing research on all that happening in AI world. So it's researching some news as well, summarizing the new updates for me. So it's just one single thing I see every morning to decide what I have to focus on today.

Utsav Somani: And it's scheduled 9 a.m. every day.

Vishesh Khurana (Shiprocket): It's scheduled every morning and it just gives me everything.

Utsav Somani: Crazy. How are we doing on Repl.it?

Vishesh Khurana (Shiprocket): Yeah. So yeah, there is some first version MVP. So it says here is members, chat, matches, calendar, admin. Yeah. So that's the V1. It's still working. It's not done yet, but you can see what it's doing.

Utsav Somani: This will be wrapped into an app so you can publish it to the iOS store. Is publishing on the iOS store easy or will you use a test flight link to bring it on?

Vishesh Khurana (Shiprocket): Developer account. It's quite easy, actually. You just need a developer account and yeah, we can publish it using the developer account.

Utsav Somani: Crazy the kind of stuff that you can do. Talk to us about the trading bot that you've built connecting to Zerodha MCV because on Twitter and LinkedIn, I mean, everyone's talking about they've built this poly market bot, how they've built this share market trading bot and connecting there. I mean, giving it $50,000 and it's making a profit, beating indexes, beating prediction markets as well.

Vishesh Khurana (Shiprocket): They actually met a lot of these kids in Miami who have built a lot of these bots on crypto mostly because there you can do a very high speed trading like second kind of trades and you can't move as fast. So it's just a new version of algo trading. For me, I built my own automation. This is the thing, how it looks like.

Utsav Somani: And how did you make this? Talk to us about the process.

Vishesh Khurana (Shiprocket): Same process here. It's basically researching and everything that's happening on X, news, basic stock analysis trigger, the standard trigger, how do you analyze the stock coming up with daily trades and it has access to my Zerodha so it can actually execute trades.

Utsav Somani: And you've set guardrails, I'm guessing for this bot where you say he can even remove.

Vishesh Khurana (Shiprocket): So I can say disable the risk manager completely.

Utsav Somani: That is insane. And tell us about the stack, like which tool did you use first to build the prompt? Was it chatGPT? I mean that same flow again? This was for your plot.

Vishesh Khurana (Shiprocket): Prompt was GPT, plot code, build the whole thing, posted on railway, which is just a hosting provider. Super based on the backend. We have just a basic stack and very, very standard. Claude decides the stack itself. So I'm not technical enough to go and decide the better stack. So I just let them do themselves.

Utsav Somani: And the difference between Claude code and doing this in regular Claude AI is basically code is slightly more technical where you can, I mean, execute like, I mean, make like proper code.

Vishesh Khurana (Shiprocket): Claude code is using Claude code as an orchestrator. So everything combined in one, so it can access your laptop files and the other files as well.

Utsav Somani: So if you were to add a feature to this, like suppose you want a live ticker, I don't see whether live ticker is right now on this.

Vishesh Khurana (Shiprocket): Yeah. So you can just do it here. So it's basically live ticker, scanner, Google news. So you wanted a live stock ticker.

Utsav Somani: I think you might need some.

Vishesh Khurana (Shiprocket): Let it do. It will add its own thing. So this is my version of connecting to it. And I can just talk to this guy and have it do it. So I've given it a prompt. Let's go back to your prompt. It's ready to publish. So this is what it's built, the offline. It's basically seen the design from offline website, try to match it. And yeah, it's basically created all the full app. This is the events calendar. I can RSVP here. It's working. It says a V1, but you can see this is what five minutes can do.

Utsav Somani: You can prompt it to like, I mean, this replete agent to prompt it to do it for iOS, do it for play store. And then it'll guide you through the process to publish it.

Utsav Somani: Alrighty. We're back this time affected by a power outage at Vishesh's end. I think all the agents are sucking up the electricity from your area. So talk to us a little bit about, I think you were building the ticker for the app, the trading bot. So it's still happening. And if something doesn't have an MCP server, like Zerodha was very early to build the MCP server and connectors out. Google is also expanding. I think Microsoft also came up with MCP connectors for Claude also. So if something doesn't have an MCP server, like WhatsApp doesn't have, like, what can we do to connect that tool with Claude?

Vishesh Khurana (Shiprocket): So for WhatsApp, like I said, I went and did my WhatsApp web APIs. So it can actually read through it. So I can connect, like you connect a WhatsApp web to your browser. There's a workaround. So there are workarounds for most of these things. For example, everything that is browser-based. Perplexity computer has access to your browser. So it can navigate within the browser. Let's say it doesn't have an MCP. So it can actually navigate using your browser and click where it's needed and do the task.

Utsav Somani: Files and data from there, basically.

Vishesh Khurana (Shiprocket): Yeah. So a lot of these tasks, browser-based tasks, you can actually have these browser extensions, like Claude for Chrome, Perplexity Comet, or Charge GPD Atlas. They can do perform tasks within the browser. So a lot of these work like that.

Utsav Somani: Where are we on the co-working? Yeah, it's working. It's going to take a while.

Vishesh Khurana (Shiprocket): Now we're expecting too much, too fast.

Utsav Somani: So let's answer some community questions. I think there are people who are asking us as well. So as a founder managing both on-ground ops and a crucial software stack, where those two worlds often move at very different speeds, what's your gut check for deciding which parts of the business actually deserve an AI agent versus which ones you leave to humans?

Vishesh Khurana (Shiprocket): See, I think every workflow, there has to be some version of agent in that workflow. Because if you are not implementing, my biggest worry would be if my people in my company are not, especially the tech and product engineers, if they are not coming and asking for a Claude's max plan or credits, I think I'll be worried. Because if I'm non-technical and my people are not asking for it, that would be a biggest one. But yeah, I would still, every workflow, let's say sales, lead farming, you know, like any human of selling guy, I think, yes, maybe that's not fully automatable. But all the things that he did previously, like creating documents, today's sales can be done in a much nicer way. If I'm a software stack, I can actually create a real-time view of how my customers, let's say D2C marketing app, I can actually create a real-time view, MVP, of showing them how my app will look.

Utsav Somani: All right. I think anything that can go wrong is going wrong today, but we'll power through. I think agents are sucking up all the energy and today the electricity gods have not blessed us. I think we're worried that the AI overlords might take over. But Vishesh, give us the updates on where the co-work is, but I'll ask you the next question. So, summarizing your response to the last one, AI agent versus human, look at your workflows as founders, as company leaders, managers, and remove as much redundancy as possible and try to automate everything and humans will still be required. So, focus on where humans cannot kill themselves and improve parts of the business and outcome as well.

Vishesh Khurana (Shiprocket): For process, especially finance and legal, where the knowledge has been already, it's a very straightforward stuff. The finance and legal, I think is the most critical ones where I can, all my SHAs reading and all where we used to pay so much can be analyzed using these tasks and a lot of money can be saved on those things.

Utsav Somani: So, the next question from Aditya of Theta Lab, I think is very related to it because it's becoming easier to build AI apps and agents today, but they fall apart in deployment. And like you mentioned that you will get an MVP, but I think sometimes you really need to take it a step further and go deeper, but it's a good place to start with MVP and you might need some technical skills to go to level two, but how do you ensure reliability when you go to real users?

Vishesh Khurana (Shiprocket): Yes, for me, I am also new. These are my first two apps. Reliability in terms of uptime and everything else, I built something where every morning, I have a recurring routine job where some agents are testing my apps like a real user every few hours. So, something and of course, I have my Sentry and all these like alarm apps, which keeps checking your apps and if some app goes down, it gives me an alert. And then from a real user perspective, for example, my app, which is the Founder Console, it's so complex, like so many things can go wrong. So many things happening here.

Utsav Somani: What is the Founder Console? And what did you use to build this?

Vishesh Khurana (Shiprocket): The same, so it was a stack, plus everything else for database, I have Superbase and it's a much more high consumption from a data point of view and integration. For me, I just built it from my own learnings of what all we used to do at Shiprocket. In terms as a founder, what matters to me, what information I need every day. It's a console where I can integrate every source of information that I get from everything I use as a company. I use payment gateways, I use accounting software, I use HR software, I use all of these things like marketing and analytics. Why not integrate all of this data into a single platform? So as a founder, I can just come in the morning, see what's happening with the company my bank balance, my runway, ARR, projections, users.

Utsav Somani: And you can give knowledge like churn and everything. So I mean, this is smart. Like, I mean, it's like a co-pilot for a founder.

Vishesh Khurana (Shiprocket): Yeah, it's a co-pilot actually. Basically, I can ask my co-pilot questions around my business. So what is the riskiest assumption that is happening in my business? And on the backend, it has skills of a Bain and a consultant mind frame. It's actually analyzing my business every second to keep me updated real time on everything that's happening. So that nothing comes as a surprise.

Utsav Somani: And this is open to all or is this in a private beta right now?

Vishesh Khurana (Shiprocket): Beta, it's open for early customers. It's a live platform. I have about a few, I think 40-50 people on it using it right now. But the reason part of this was...

Utsav Somani: Founderconsole.ai. Folks log in and create your account, give us your feedback.

Vishesh Khurana (Shiprocket): Please do. I would love that. And then I have built a small simulator. This is basically running multiple scenarios. AI can do multiple scenarios and give me on each of the scenarios. Let's say if my payroll reduces by 20%, what happens? If my marketing cost goes up by 20%, what happens? And then generate a strategy document for me, which includes all of the information that it has and gives me an advice on how, what I should do. But this is a demo company, of course.

Utsav Somani: You made this in cloud or have you used other tools as well?

Vishesh Khurana (Shiprocket): Replit and for logics and all I have to, I again fall back to open AI and stuff. So each of these like simulator was a very critical piece. A lot of calculations and a lot of number crunching is happening on the backend. A lot of the content work in terms of research for web happening. So everything is linked, like all my APIs. Even the co-pilot has access to all the elements on the backend. Time taken to build this, cost to run it monthly? Time, I think it took me about 15-20 days to put everything together. Monthly, there's not a lot of traffic on it. So it's sitting on a $50 server. But of course, I have to scale it as it goes further.

Utsav Somani: And right now you're not charging for it? Like it's open to all?

Vishesh Khurana (Shiprocket): It's open, it's free. I'm not charging anything. It's basically free.

Utsav Somani: And you can connect like Google Sheets?

Vishesh Khurana (Shiprocket): I have integrations of everything. So Google Analytics, GA, Salesforce, HubSpot, everything that we possibly can use. And if you are using something that you need to integrate, I can put that. HR softwares, Google Sheets, you asked for. So everything comes into a single platform. And for example, this is the ship rocket. So it's now creating, analyzing all my data and creating a strategy document for me. But this was a very heavy, if you look at the amount of backend tasks. So obviously, I had to split this as advised by again, AI. So split it between multiple stacks. Database has to host it somewhere else. For example, for it to move as fast as it can. There's a lot of backend parallel tasks that are happening together.

Vishesh Khurana (Shiprocket): But it comes up like, yeah, that was one of them. So for this, how do I keep it up? There's so many features and everything else. So I have agents going through, logging into the platform every morning as a user, checking everything. And if something goes wrong, give me a rip. But that's just my small version of doing this.

Utsav Somani: Wow. Wow. So where are we on the cloud? This thing now?

Vishesh Khurana (Shiprocket): It should have done its job. Livestock trigger, dashboard, code pushed. I have to just deploy. It's done. Let's see what it's made. You have to give it five minutes for it to redeploy and just show up and propagate across the internet. Yeah, it's done. Its coding job is done. It's created what we asked for. It's just redeployed, triggered. It's going to take a second.

Utsav Somani: Another question from Ishan. In product management, the 101 rule is to always think about building product from the perspective of your laziest user. How does this inform the product you're building and the subsequent design and the UI UX decisions you make?

Vishesh Khurana (Shiprocket): I've just been very lazy. Personally, I've just used agent to go through it, log on to it like a real user every day and just reiterate because I can do this a hundred times now. It's least cost to me. So one, I have an agent that is now designed as a master UI UX researcher who's researching the best products in the world, how the UI UX is, and then taking that knowledge and seeing that my product and trying to marry both. So I'm very lazy, just asking the agents every question and let him decide the best version of it.

Utsav Somani: I was redoing my personal website on cloud and I think every output in cloud, I think that's the default thing. So today I actually downloaded a skill to add new designs. GitHub exactly.

Vishesh Khurana (Shiprocket): Download some repos and there are so many people who've done amazing job on GitHub, ScreamerMotion and all of these things to create beautiful looking websites which don't look bicoded now. Because bi-coded websites have started looking the same.

Utsav Somani: All looking the same and even, I mean, there are certain tells, right? Shorter sentences, em dashes on OpenAI, Claude made that beige color which comes automatically. So you can feed in like different things. So yeah.

Utsav Somani: It's coming. I think connect ticker data, I think you'll have to connect kite first.

Vishesh Khurana (Shiprocket): Oh yeah, I have to do my login again.

Utsav Somani: I think refresh because I think it's ticking the ticker data, connect kite first.

Vishesh Khurana (Shiprocket): That's connected. This was what I asked, news.

Vishesh Khurana (Shiprocket): Livestock dashboard header, it's on the header.

Utsav Somani: Also this, I think, okay, market closed, yeah.

Vishesh Khurana (Shiprocket): Market closed, that's why it's not coming up. Because live is live.

Utsav Somani: Interesting. So I mean, basically, I mean, your curiosity is the blocker. But I mean, how should founders think about like zooming out and becoming more philosophical now? So you've, thanks for sharing all of this, coming on our show as well and doing our first masterclass. But how should founders think about their existing companies? If you were building Shiprocket now, what would you do differently in the world of AI, given that you know everything about AI tools that you've experienced personally?

Vishesh Khurana (Shiprocket): So I think the first step everyone should do is go back and analyze the human capital. What each and every individual is doing in the company? And are those individuals now evolved and using the new stack to do their job? So if there are people who have not evolved to the new world of AI, if someone today tells me, oh, AI is not capable, that's the first guy I would fire. So one, everyone should adopt the new world. So you look at every individual, are they using this and are they evolving and adapting to the new world and are getting more productive? That's definitely going to be the most critical first step. Because in all of our companies, the highest cost is the human capital. And that's where the most margin lies. So that's going to be the first step one. I would rather, and if they are not, I would spend a lot of time to train a lot of these people, give them access to tools, get enterprise accounts, give them these tools, make sure they're using all of it to be more productive. And then of course, over time, it's unavoidable. We can mince our words and play around it. But the reality is, there will be job cuts and a lot of things will be taken over. And who does faster? Now that's the job. You might not feel good about it. You might feel all of those responsibilities on your head. But that's the reality. If you might not do it with someone, your competition, someone will do it. You'll have to now live with this.

Utsav Somani: I mean, disrupted by an AI native firm doing the exact same thing, but running with a limited set of people as well, right? You've said that 500 million, billion dollar outcomes are possible with teams of less than 10.

Vishesh Khurana (Shiprocket): Or one, there's one guy, Mediwee some company, I had to go and search. This company was just a mid-tech guy, one single guy, billion dollars. That's almost less than 20 people, never raised any external funding.

Utsav Somani: So that is insane. That is insane. So I mean, yeah, so that is what I think is exciting. But what is keeping you busy? What are some of the new tools that you personally used in the last few days or last few weeks that you think you want to go deeper on?

Vishesh Khurana (Shiprocket): I am definitely going deeper on Kimi, the agent swarm style, how it operates. That's one thing I'm doing very, trying to get like the whole micro SaaS idea. So I am trying to come up with this micro SaaS. Clear will be first in the list, which I want to launch where I want individual distribution becoming the key. So now I'm trying to figure out, it's not about the idea, it's not about the tech stack, that's all commoditized. So it's not about my tech is better than yours. Everyone has a better tech now.

Utsav Somani: Come and sponsor on offline network. I think that's the best way to earn distribution or clips do really, really well.

Vishesh Khurana (Shiprocket): All about distribution. That's all about it.

Utsav Somani: But you're excited to continue building. There was one tool that you were mentioning before the call, I think which was exciting. Some new thing that you were trying out?

Vishesh Khurana (Shiprocket): Right now I'm obviously spending most of what I'm in Founder Console and Clear, which are the two tools I want to now distribute. The problem with this also is, it's just so addictive and exciting that you keep building new and new things and forget and build and forget. A lot of you, I met Sahil the other day, Shiprocket. He has built so many tools, like randomly and all of them are sitting somewhere hosted, live, but just forgot.

Utsav Somani: Live coding is the new video gaming, I think.

Vishesh Khurana (Shiprocket): People are having fun with it. I am now trying to go back and discipline myself saying that stop building and let's first make sure these tools are actually useful. And just one interesting thing I am very curious about is called predict Amiro fish. Let me go back and share my screen one more time. It's a prediction software. I'm trying to make it better, but very interesting tool. Can you see my screen?

Vishesh Khurana (Shiprocket): You know what? It's a GitHub repo. I just picked it up and hosted it on my white tablet for myself. So you can give it a document. Let's say I have a document on what's happening with the AI in the world. I can just give this document to this tool. I hope this works. Normally, everything that works falls in live demos. So now, what will, this is something I'm very excited about. So what it does is, it synthetically, I've given it a world event, what is happening and asked it to predict future of this event.

Utsav Somani: I mean, do, I mean, predict so many different things, like future of relationships, future of your personal finances, future of your company, like, I mean, opens up a can of worms.

Vishesh Khurana (Shiprocket): Exactly. So this tool, it's a very, it's just the beginning of this whole thing. What it does is it creates synthetic humans, like thousands and thousands of people. So you know what it does, what it did right now? It created an AI researcher, tech journalist, person, AI user, community, business entity. It's creating multiple humans and synthetic feelings. What will happen to predict how humans will respond to this event? So for wars, how will humans respond? So last year, I did this for Israel and US, this Iran war. So it generated Donald Trump, a US president, Israel's president, Iran's president, how and what they will do as eventually. So now you see this on the left.

Utsav Somani: I would love to have a poly market bot on this, which is betting on prediction markets across the world. Empowered by this, actually.

Vishesh Khurana (Shiprocket): Yeah, this is a very early stage. I'm trying to just, I'm doing a lot of my time just trying to play around with this and how to make it better and make it more smarter. But this was one of the most interesting tools I came across, just to have fun. So it's creating these nodes of all the related parties in this universe. For AI, you see what's happening? Yeah, it's created first open AI is Gemini, the main then people, what they do, it's going to keep expanding. So if we spend this time, five, seven minutes on this globe, this is actually consuming over 200,000 tokens right now.

Utsav Somani: And it gives you probabilistic answers at the end, like probability of this happening, probability of this happening, I need a percentage.

Vishesh Khurana (Shiprocket): What will happen? Very interesting tool. I just like to play around with it.

Utsav Somani: You can upload your company finances, give it context about your company, the things that you're doing. I think that's one thing that founders can play around with.

Vishesh Khurana (Shiprocket): I'm building this within founder console. So I will just add this into founder console itself.

Utsav Somani: Nice. All right. I think we're over time. So thank you so much for coming on. And we'd love to have you back also, as you keep exploring these tools. We'll do a version two of this masterclass also, hopefully with better internet and electricity as well.

Vishesh Khurana (Shiprocket): I'll come to you. I'll sit with you next time.

Utsav Somani: We'll do it in person together. That makes sense. Our book is out today, by the way, The Founder Manual. It's available on Amazon and Flipkart. If you want to read Vishesh's backstory on how he discovered the idea for Shiprocket, evolved from Cartrocket in a pickle store, I think, while buying pickles. Grab the book and give us your feedback and there you'll know the Shiprocket journey and many other founders who openly shared with us from offline. Thank you so much. And Vishesh, thank you for coming on the show.