I can’t believe we built an AI employee in 62 mins (Cursor, ChatGPT, Gibson)

1. 0:00 – 3:41

   Build plan + tech stack for an AI customer success “employee”

   1. HMHarish Mukhami0:00

      Today, we'll be building an AI customer success agent, which is gonna solve problems for you without any humans involved. We're gonna be using Gibson AI database for database backend. We'll be using Cursor on Claude Sonnet 3.7 for coding. And finally, we'll be using CrewAI framework for AI agents. We're building this in three parts. So part one, we'll build a SaaS tile dashboard app to provide insights and recommend actions. Part two, we'll add AI agents to analyze these insights and recommend actions. Part three, we'll actually make this agent autonomous and take actions on users' behalf. This is the dashboard. This is a result of part one. This is the customer success AI agent dashboard only, so we're focusing on all the way from design, development, deployment, management, and scaling of databases, which coding tools don't do today. So when we started building in Gibson, we used a web interface. Now we're actually building tables right from Cursor IDE a-and deploying the database. You're not building a prototype anymore, so you are backed by a very significant production-grade database. It's already deployed in the cloud, so you publish this, and you get 10,000 users tomorrow.

   2. AGAakash Gupta1:05

      What other AI employees could people create?

   3. HMHarish Mukhami1:08

      SDRs are a very popular use case. There is also meeting preparation agent, executive assistant. You can chain together all of these agents to, to create a specific role that is tailored for you.

   4. AGAakash Gupta1:22

      Really quickly, I think a crazy stat is that more than fifty percent of you listening are not subscribed. If you can subscribe on YouTube, follow on Apple or Spotify podcasts, my commitment to you is that we'll continue to make this content better and better. And now on to today's episode. All right, Harish, welcome to the podcast. What are we gonna be doing today?

   5. HMHarish Mukhami1:46

      Yeah. Today we'll be building an AI customer success agent, uh, which is gonna solve problems for you without any humans involved.

   6. AGAakash Gupta1:53

      All right. Let's just get into it.

   7. HMHarish Mukhami1:56

      Yeah. Uh, let me share my screen. So we're gonna be using four tools today. Uh, we're gonna be using OpenAI o3-mini for planning. We're gonna be using Gibson AI database for database backend. Um, I can do a quick intro of what Gibson is. Gibson is a, a AI-powered cloud database. You can build, deploy, and manage your database at lightning speed. Uh, so it's, it's, it forms the database backend. And we'll be using Cursor, uh, on Claude Sonnet 3.7 for coding. And finally, we'll be using CrewAI framework for AI agents. So that's the tech stack that we'll be using today. Um, so, um, we can, we can go on and start building.

   8. AGAakash Gupta2:37

      Let's do it.

   9. HMHarish Mukhami2:40

      So how are we building this? Um, we're building this in three parts. Uh, so part one, we'll build a SaaS tile, uh, dashboard app to provide insights and recommend actions. So this is something that most of us are very familiar with, so which is a dashboard with metrics and insights, which we have been using for a while now. Uh, part two, we'll add AI agents to analyze these insights and recommend actions, essentially automating out, uh, the first part. So now AI agents will be looking at the dashboard, looking at the data, and analyzing, um, the data, uh, identifying churn, uh, and mitigation reasons, and providing actions to users who can approve or take those actions. So this is like a part two is the human in the loop, uh, with an AI agent. Uh, part three, we'll actually make this agent autonomous and take actions on users' behalf. So at the end of it, we'll have a fully functional AI agent for customer success. So let's dive into part one. Uh, so for part one, uh, we'll start

2. 3:41 – 5:42

   Using o3-mini to design the dashboard tabs and success metrics

   1. HMHarish Mukhami3:41

      with planning phase. So we know that we wanna build a dashboard for AI customer success agent, how exactly dashboards should be, what metrics should we focus on. So we're gonna ask, uh, uh, ChatGPT, in this case o3-mini, to provide us some guidance here. So I have a quick prompt ready to go. Uh, so I'm gonna copy-paste this here. Um, fairly straightforward prompt. "Act as a business intelligence expert to design a comprehensive dashboard for Bulk Trade," uh, that is, that is a company that we're focusing on, "a B2B e-commerce platform." Uh, you can imagine this could be for any other business, but today we're, we're gonna focus on the B2B e-commerce. Uh, "As a customer success manager, I need dashboard that integrates data from our CRM," uh, in this case we're using HubSpot, "Google Analytics, or an application database. Uh, the goal is to empower our customer success team with actionable insights that will optimize customer funnel, improve engagement, retention, reduce churn. Uh, please provide a dashboard with four tabs, sidebar needed to ensure that we achieve our goals." Um, I'm spelling this out because this is the one that kicks off everything in motion.

   2. AGAakash Gupta4:47

      Okay.

   3. HMHarish Mukhami4:47

      So I'm gonna ask this, uh, and wait for, um, or using it in reasoning mode, so it came back with pretty, pretty quickly came back with a few, few, um, recommendations here. So, uh, funnel performance, engagement and usage, customer health risk, um, retention and churn. So four tabs like we asked, uh, based on what we, um, what the goals are here.

   4. AGAakash Gupta5:14

      So why did we use o3-mini reasoning mode?

   5. HMHarish Mukhami5:18

      So, um, of all the models that I've f- uh, that I've used, I've found o3-mini is great for planning, um, especially it's much faster. Um, so, uh, it has worked out well for me so far. For coding, I use Claude Sonnet. That has been the king, I guess, for the past year or so. So those are the, the two models that we're gonna be using today.

   6. AGAakash Gupta5:39

      Sounds good.

3. 5:42 – 8:16

   Translating dashboard needs into a data model (fields + sources)

   1. HMHarish Mukhami5:42

      Um, now we, we have the dashboard design. Uh, the second and, and even more important thing to focus on is what is the makeup of your database that is powering your dashboard? So I'm gonna just ask o3 to see, um, uh, to sh- tell us, like, what data should we use? Um, what are some of the components here? So what data powers this dashboard?

   2. AGAakash Gupta6:09

      So now it's thinking about it and helping us plan what data inputs we'll need to create.

   3. HMHarish Mukhami6:15

      Exactly. Um, so it, it is funnel performance, engagement, and usage. So one of the things it just did was it did exactly like the dashboard organization. So it gave, for each dashboard, um, these are all the data fields that you need. What I do wanna ask is, like, organize this by data source instead of tabs.

   4. AGAakash Gupta6:43

      Makes sense.

   5. HMHarish Mukhami6:50

      So now we have a much more specific, um, response around like, "Hey, from HubSpot, these are the, these are the fields that you need." I'm gonna copy-paste this whole thing.

   6. AGAakash Gupta7:02

      It's very much like a data definition.

   7. HMHarish Mukhami7:06

      Exactly. You don't even-- I, I guess you don't even have to have this specific specificity, uh, um, with, uh, from a database perspective. Uh, you got a bunch of these. So I'm gonna copy this, and then, uh, we'll go to Gibson and model this database.

   8. AGAakash Gupta7:27

      Okay. And so what role is Gibson playing here?

   9. HMHarish Mukhami7:32

      So Gibson is the database that is powering this entire application. So, uh, first step when we're building the dashboard, um, the front end is actually looking into the Gibson AI database to, uh, to show that data in the front of the dashboard, in the dashboard. One of the cool things about Gibson is it's an AI-powered cloud database. So you could imagine, um, if you wanna build this analytical back end for, for like, um, um, customer success dashboard, you need to have a database that is, um, easily scalable, can, can ingest data from multiple sources and handle, like, hundreds of thousands of transactions, if not millions. Uh, Gibson can autoscale with that, so that's why I picked this.

   10. AGAakash Gupta8:14

       Okay.

4. 8:16 – 12:43

   Creating a production-grade database in Gibson (web workflow)

   1. HMHarish Mukhami8:16

      So let's go to Gibson. So Gibson is available on, on the web, or it's also available in your favorite tools via MCP server. Um, so we're gonna, we're gonna see both. So I'm gonna start off with the web, uh, version of the Gibson AI platform. So I'm just gonna sign in. It's free to get started. Anybody can join. You can use your Google Auth or log in with your email. I'm just gonna use my email to join in. And then this is the Gibson AI workspace. The first thing I'm gonna do is, uh, type in "Build me a database for B2B e-commerce customer success agent." So I'm just gonna type and, and provide a simple prompt. This is not the actual prompt that is gonna build the database but initializes the project and, uh, sets the context on what the user wants to build.

   2. AGAakash Gupta9:19

      That's our high-level goal, and then we're gonna come in with the data definition later, I assume, from O3 Mini.

   3. HMHarish Mukhami9:24

      Yep. Yep, absolutely. So, uh, the one we copy-pasted, I massaged this a little bit to make sure that, like, it's a little bit more cleaner. Um, we're still gonna have those three components to it. Um, here is your project window. So these are all the projects that you're working on. They show up on the left side. Uh, so the first step is it initialized your B2C-- EB-- B2B ecommerce customer success agent, um, and it gives you two options. I can either design and generate the entire data model for you at once, or you can build the data model step by step. So you would choose an option one if you weren't using O3 Mini, you're just asking it to, like... And you can work with it to design and brainstorm. Since we have already planned in O3 Mini, so I'm gonna go choose the option two here.

   4. AGAakash Gupta10:13

      Today's episode is brought to you by Amplitude. Building great digital products is hard. You know that better than anyone. Getting teams aligned, measuring what matters, and scaling your product strategy isn't easy. But what if you had a clear framework to guide your next steps? That's exactly what Amplitude built. They studied the best product teams to understand what really drives impact and turn those insights into the digital experience maturity assessment. In two minutes, you'll be able to see where your team stands and what you can improve to build better products faster. Click the link in the caption to take the free assessment and get a clear path to product growth. Hey, let me take a quick break to talk about something that's completely changed my product management workflow: Linear. As a PM, I was drowning in tools, one for planning, another for issue tracking, roadmaps and sheets, and jumping between Slack, Intercom, and app reviews just to piece together customer feedback. Sound familiar? I was spending more time keeping systems in sync than actually building product. Every time development kicked off, my carefully crafted plans would immediately need updating. I was the human API between all our teams, constantly chasing updates and translating between tools. That's why I love Linear. I can capture customer feedback, shape product ideas collaboratively, quarterback cross-functional teams, and monitor development progress in one place. It cuts through the maze of disconnected systems that were complicating my life. Product teams at OpenAI, Vercel, Brex, and Cash App all use Linear. If you're tired of spending your days keeping different tools and teams in sync, check out Linear at linear.app/partners/aakash. That's linear.app/partners/aakash. So option two says we can choose-- build the data model step by step as we see fit.

   5. HMHarish Mukhami11:45

      E-exactly. And here, um, the goal is we're more s- uh, more d- uh, directional and targeted about what we want Gibson to build. It does give you some options, but since we already have it, so I'm just gonna copy-paste that. So...

   6. AGAakash Gupta12:03

      And do you have several models working in the back end here? What is the-- What is powering this experience?

   7. HMHarish Mukhami12:09

      Yep. We use both Anthropic as well as, um, OpenAI.Uh, so we, um, actually use Anthropic for writing the code and OpenAI for actually planning and infrastructure management. So both of those are available. This is the crux of what Gibson does since we're building data model from scratch. So this takes about, um, a minute or so because it's a pretty significant prompt. So while it is building, I actually wanna walk through pretty quickly what are all are there on the workspace. Um-

   8. AGAakash Gupta12:41

      Sounds good.

5. 12:43 – 17:23

   Schema review, deployment, and instant CRUD APIs

   1. HMHarish Mukhami12:43

      Uh, so while it is doing, while it is doi- uh, doing the, like the data model and schema design, um, when Gibson is done, you'll get a complete data model schema back, and then it is represented in an ER diagram here. That's the first thing. The second would be a schema. So schema would be your, um, um, Python code, um, in a schema format. Uh, and then if you like that da- data model that you have designed with Gibson, uh, and then if you wanna deploy it, you just hit Deploy. So when you hit Deploy, uh, we're gonna take this, uh, schema and data model and deploy it in our AWS instance. It's a secure, scalable, uh, cloud infrastructure that we have made it available for users to use. So at the end of it, so it's done. It's, uh, it built out what you asked, and then it comes back with like a, a pretty specific response around like, "Hey, I built these following three modules: CRM, analytics, transactional data module." Um, so let's go quickly look into, um... This is a ER diagram, so on the CRM side, um, contact information, feedback, profile. So very cleanly written out with, uh, with the data types, ID is your key, and then it has some annotations around what the data model is and what are some of the constraints here. Uh-

   2. AGAakash Gupta14:05

      Very cool

   3. HMHarish Mukhami14:06

      ... second is the schema. Uh, so like I mentioned, like schema is actually like the code, uh, representation of what the data model is. Uh, from here on out, like you can actually ask, uh, Gibson to do anything. It can explain you anything. It can pro-- it can be, it can be your like, uh, a pair data model or database agent. So one of the things that we're gonna be doing is we're gonna be ingesting some test data into it. So what I'm gonna ask Gibson to do is to remove all the enums because it becomes challenging to insert test data. So please remove enums from all the tables.

   4. AGAakash Gupta14:44

      And for people who don't know, what are enums?

   5. HMHarish Mukhami14:47

      Uh, enums are enumerator, enumeration. So when you are defining fields, uh, on a table, there comes many different data types. Um, enums is one of the database data types and which locks in very specific values like zero, one, two, three. For example, if a customer is, uh, a lead, um, a, a customer or churned or any of those, those becomes like zero, one, two, three. So, oh, there are no enums data avail- uh, so all of them are text, so we don't have to worry about that. So...

   6. AGAakash Gupta15:20

      All right.

   7. HMHarish Mukhami15:20

      So typically I just run that, but we don't have any enums in this data model, so we're good.

   8. AGAakash Gupta15:24

      Okay.

   9. HMHarish Mukhami15:27

      Um, couple other things. So once we deploy, um, I'm just gonna hit Deploy here. Um, it's gonna go take this data model and deploy it in, uh, in, in the cloud. And then as soon it is done, it's gonna give you two environments. One is development environment, another is, uh, um, production environment. So automatically provisions two environments for you to, uh, develop, test, and then if you like your data model, then you can, you can promote it to production. So, uh, click on, uh, View APIs. So it's gonna show you, um, your entire APIs that are backed by this database. So within under three minutes you have this entire CRUD layer available for you. So, um, you got the entire database, uh, access through this CRUD layer. So this is your API-

   10. AGAakash Gupta16:19

       And a CRUD layer is?

   11. HMHarish Mukhami16:21

       Uh, create, uh, read, update, delete. So those are the operations that you can perform in the database via the, um, via the-

   12. AGAakash Gupta16:30

       Get, post, delete type commands. Okay.

   13. HMHarish Mukhami16:32

       Exactly. I'm gonna quickly show you the settings. Uh, in Settings, um, you have both the dev and production, and these are the API keys. I'm gonna delete this later, so, uh, it's okay to-

   14. AGAakash Gupta16:48

       Who's watching? [laughs]

   15. HMHarish Mukhami16:51

       Uh, and then, uh, and then if you wanna direct access this, so if you don't wanna use the database and you, uh, don't wanna use the APIs, you, you also have the opportunity to directly take this connection string, copy-paste it, and then access the database here. So we have the data model built out. So, um, what we can do now, now that we have the data model, um, let me see if there's anything else I wanna show. I think we're good. Uh, we can go to Cursor, and then we can ask Cursor to build a dashboard based on this data model.

   16. AGAakash Gupta17:21

       Exciting.

6. 17:23 – 23:39

   Connecting Gibson to Cursor via MCP (database operations inside the IDE)

   1. HMHarish Mukhami17:23

      So let me share my, uh, Cursor window here. So here we are in Cursor. Uh, so the first thing you see is how we have connected Gibson to Cursor. Um, I've-- uh, Gibson connects to Cursor or any other dev tool via MCP. Um, what is MCP real quick? MCP stands for Model Context Protocol. Um, it's an open standard designed to connect AI models to, uh, your tools, your data sources. Um, MCP makes it easy, um, for tools to integrate with LLMs and vice versa. So this has been like a really game changer for us, like even while we're doing this evolution with Gibson. Um, now with MCP, it's so seamless. Everything is so seamless, uh, on how we interact. Um, how do you integrate with, uh, integ- how do you integrate Gibson with Cursor? It's a very simple command. So just four lines, Gibson is in Cursor now. Um, what I'm gonna do is I'm gonna just go and show you, um-I'm wondering, like, what can I do in Gibson from, from Cursor? What are some of the tools available? So I'm just gonna ask, like, what tools are available via Gibson MCP server? So it's gonna tell us, like, what are all the capabilities that we have available in the IDE. Um, the cool thing about this is developers don't have to leave their environment, and without leaving, they can actually do, uh, project management, create, get, update projects, schema management, get all the schemas, do the data modeling, deploy the databases. Um, and one, uh, this is my personal favorite. You can actually run natural language queries on your database from here, and it can, it can return the results. So, uh, literally text to SQL with, like, um, curated results back to you. So those are all the tools that are available here. Uh, first thing we're gonna do, since our database doesn't have data yet, is I'm gonna ask Cursor to generate test data. Um, please generate test data. Um, oh, I need to do one quick thing. So I need to rename the database so that it's not confusing, because since we have many projects here.

   2. AGAakash Gupta19:39

      Okay.

   3. HMHarish Mukhami19:40

      Uh, I'll, I'll just do it in the Gibson end. So I've just renamed the project here to, um, CSA DB one, so that it's... Since I have plenty of projects here. So let me quickly make sure that took into effect, and then I will ask Cursor to generate test data. There are a couple things I said. I'm very specific about using API calls and then also honor relationships, uh, between the underlying tables, because we have a bunch of tables here. So I'm just gonna say use APIs and specifically create a Python script for this.

   4. AGAakash Gupta20:20

      Okay.

   5. HMHarish Mukhami20:23

      So if you f- if you see here, so it's actually going to Gibson and then getting the project information. So this is how Cursor has the context of whatever we have built in some third-party infrastructure.

   6. AGAakash Gupta20:36

      Got it. And as far as the MCP connection that we were covering, what was the steps involved to do that?

   7. HMHarish Mukhami20:43

      Um, so first step is actually you need to authenticate that you're a, a authenticated user of Gibson. So we have a blog post out there on, on our website that shows, like, it's what first step. Um, the second would be just taking those four lines, um, going to Cursor settings, clicking here, and then this opens up this tab, and then you go to MCP, and then you just add a new MCP server.

   8. AGAakash Gupta21:07

      Okay.

   9. HMHarish Mukhami21:10

      So if you see on the, on the, on the conversation window, Cursor is actually getting all the API information. So now it knows the context. Now it can go create, uh, a Python script for us. Uh, so it's gonna go generate some test data for, uh, for the dashboard.

   10. AGAakash Gupta21:27

       Okay. And behind the scenes, Cursor is probably thinking about what might be good test data to make this dashboard look good. Then it's going and adding that into the databases that were provisioned by Gibson.

   11. HMHarish Mukhami21:42

       Absolutely. Yeah.

   12. AGAakash Gupta21:43

       Okay.

   13. HMHarish Mukhami21:44

       Um, I think, um, one of the key things we should probably make sure that when we do test data insertion is ma-make sure that we've mentioned realistic. Otherwise, it just pushes in some data that is, that is not necessarily doesn't make sense.

   14. AGAakash Gupta22:01

       Okay.

   15. HMHarish Mukhami22:04

       It, it does take, uh, take a, a couple seconds to do this.

   16. AGAakash Gupta22:08

       Okay. So here at the moment, Cursor is building the test data?

   17. HMHarish Mukhami22:15

       Yeah. Building the script for the test data. So we're, we're gonna run that script. Um, Gibson also has the ability to generate test data. Uh, so you-- we can just ask the query tool, I'm like, "Hey, go generate test data." But I wanted to create a Python script because it's, uh, repeatable, and then we can execute it pretty quickly.

   18. AGAakash Gupta22:32

       Okay. Got it.

   19. HMHarish Mukhami22:35

       While it's doing that, I also wanna show you a couple things. One is, um, the model selection. So I'm using the Claude Sonnet 3.7. And then the second is I'm also using thinking mode. So for some reason, I found the thinking mode delivers better results. Uh, so couple things I found while, uh, while working with Cursor.

   20. AGAakash Gupta22:56

       Okay. And what is the-- is that probably the most expensive model you can use? [chuckles]

   21. HMHarish Mukhami23:03

       I think the 3.7 Max is. I, I-- and it's actually an extra cost to it. Um, I know, I mean, nowadays, I think it was 4.1, GPT four, um, 4, 4.1, OpenAI 4.1 sounds like that seems to be pretty good at coding as well, so-

   22. AGAakash Gupta23:22

       Mm-hmm

   23. HMHarish Mukhami23:22

       ... the foundational model coding race is on. [chuckles]

   24. AGAakash Gupta23:26

       Yeah. By the time this episode airs, I'm sure somebody will have released another one.

   25. HMHarish Mukhami23:33

       I, I think there's a, a new model that's coming out today, so.

   26. AGAakash Gupta23:37

       Okay.

7. 23:39 – 25:12

   Generating and inserting realistic test data (Python + Faker)

   1. HMHarish Mukhami23:39

      Anyway, so we're done. Um, comprehensive Python, Python script that generates realistic text data, so data is ready. So in the interest of time, let's go ahead, accept it, and then look at the README. What does it say? Okay, we have to install Faker. So let me do that quickly. Oh.

   2. AGAakash Gupta24:05

      And how did you know we needed to install something?

   3. HMHarish Mukhami24:08

      Um, I went to the README. So Cursor generated a README. Uh, I clicked on the README file, and then, um, it says these are, these are the couple steps that you have to use to run this, so.

   4. AGAakash Gupta24:22

      And Cursor, unlike the AI prototyping tools that a lot of people are u- probably used to, this isn't happening on a virtual machine. This is happening on your machine. So you installed that thing on your machine.

   5. HMHarish Mukhami24:31

      Absolutely. Though there's an option for you to create a virtual environment, um, so UVV and V, and we're gonna do that too today, uh, when we're running the AI agents. Uh, and it's gonna go. So now it's using that Gibson API to generate these, um, data as you see it go, go through the terminal. Uh, but you're right, I think C- Cursor does that. Uh, I definitely prefer Cursor because it gives me more control, uh, over what I'm building. Uh, though you are responsible for infrastructure, uh, that's why Gibson takes care of the back end here.

   6. AGAakash Gupta25:09

      Okay.

8. 25:12 – 31:17

   Vibe-coding the Next.js dashboard with Cursor Agent Mode

   1. HMHarish Mukhami25:12

      Uh, so it'll go through and, and insert that data. What we're gonna do next is now that Cursor has inserted, um, f- um, test data, I'm gonna go ahead and I'm gonna say, um, "Now, now that we have the data, let's go build a dashboard." And I'll, I'll walk through the prompt here. Uh, I'm being very specific on what packages to use. So build a modern, sleek and interactive dashboard based on the Gibson AI project. Uh, use the APIs to power the dashboard. Uh, don't use mock data. So sometimes Cursor ends up showing a dash- dashboard but there's mac- mock data behind it. Um, get API information. So I'm t- uh, and then I'm being very specific about use Next.js and ChatCN for UI components for dashboarding. So that's the instruction. What to build, I'm gonna go to the ChatGPT on what it has given us from a guidance perspective for the dashboard and just copy-paste it just to keep things sim- simple here. So, so I copy this and then I'm gonna go copy-paste this from GPT. So this is... Hang on. This is exactly... We got... And then... So if you look at this, um, it's essentially... Let me, let me actually share the screen as well so that you can see where I'm copying this from. So I just took and copied this whole thing. Um, we asked a question and then we got the dashboard, and then we're just gonna copy-paste this into Cursor here. Okay. It's still, it, it's definitely creating a pretty, uh, hefty data set for us.

   2. AGAakash Gupta26:57

      So at the moment the Python script is inserting the data?

   3. HMHarish Mukhami27:00

      Yep.

   4. AGAakash Gupta27:01

      Okay.

   5. HMHarish Mukhami27:03

      We can, we can open a new... Actually we can open a new Cursor window while it is doing it, uh, so that we can do this in parallel.

   6. AGAakash Gupta27:12

      Okay.

   7. HMHarish Mukhami27:13

      So let me open up a new, new window here. Let's, let's start a fresh project here, and then I'm just gonna do this. So I'm gonna... It's more clear to read. Um, so what we talked about below is the overview of the dashboard and there's four primary site, site tabs. So I'm just gonna do a send. So since it's a pretty significant prompt, um, there, there are a few things that Cursor is gonna do since we are using Claude 3.7 with, in a thinking mode. Um, it reasons pretty good so it's... While it, while we're vibe coding we can actually watch how it's thinking.

   8. AGAakash Gupta27:50

      Uh-huh.

   9. HMHarish Mukhami27:50

      Um, I, I like to do that and maybe learn a thing or two while it's doing it.

   10. AGAakash Gupta27:56

       Yeah. It's cool that it shows you-

   11. HMHarish Mukhami27:59

       One of the things that we, um, it asks here is, um, install- installation of some packages here. So would you like your code inside a, a source directory? Yes. Would you like a Turbo pack? Yes. Uh, customize the import alias? No. And then it goes, the first step is installing the dependencies for Next.js. I'm using Next.js with ChatCN because it looks nice. We can just use Python. It's much faster but it may not be visually appealing.

   12. AGAakash Gupta28:28

       Okay.

   13. HMHarish Mukhami28:29

       This whole thing takes about 10, 15 minutes. I obviously have, um, a completed project that we can show because it ta- it does take a bit, um, to c- completely do this, so.

   14. AGAakash Gupta28:41

       Okay.

   15. HMHarish Mukhami28:42

       We'll, we'll go through a few steps and once it's running we can go and, and look at the finished, um, the first step finished product.

   16. AGAakash Gupta28:48

       Sounds good.

   17. HMHarish Mukhami28:50

       We're still at the installation phase, package installation phase.

   18. AGAakash Gupta28:53

       [laughs] Getting set up-

   19. HMHarish Mukhami28:55

       So it's done now. Now it's creating the project structure.

   20. AGAakash Gupta28:59

       Okay.

   21. HMHarish Mukhami29:00

       So this is the step where it's creating the sidebar components. By the way, we're using the agent mode here. Cursor has a few, um, multiple modes, so ask is more, um, act as a copilot agent is agentic workflows where it goes through and then builds this, um, builds a product to the final prompt and then automatically resolves those errors and issues as they come up.

   22. AGAakash Gupta29:24

       Do you normally spend most of your time in agent mode?

   23. HMHarish Mukhami29:29

       I use, uh, ask mode quite a bit for the purpose of one-shotting, uh, quick apps. Like, uh, vibe coding is, agent mode is, works pretty well. So I've used all the other tools too like Replit and Bold and, um, Lovable and all of that. Like, Cursor is the thing that I, I like the most so far.

   24. AGAakash Gupta29:52

       It's the most powerful.

   25. HMHarish Mukhami29:54

       Yeah. It is the most powerful. So let's go look at once it executes I can actually show the end state. In the interest of time we got a, we got a few other things to show through so it, it does take a while to go through this.

   26. AGAakash Gupta30:11

       Okay. Today's episode is brought to you by Maven. The problem with most courses online, like Udemy, is there's no live component and the instructors aren't experts in their fields, they are professors. At Maven you get direct live access to experts and operators from the world's best tech companies. You can't get that access anywhere else, in any university, and you usually can't find them on YouTube either.

   27. AGAakash Gupta30:33

       I featured so many of Maven's experts in the newsletter and podcast for that reason. To help you out, I've put together a collection of courses I recommend at maven.com/x/aakash. This includes courses like AI prototyping for PMs, product sense for PMs, and getting an AI PM certification. Visit it now at M-A-V-E-N.com/x/A-A-K-A-S-H. We're putting it in the oven, and magically it's gonna come out of the oven baked. [chuckles]

   28. HMHarish Mukhami31:07

       Yeah. Uh, let's do this. Uh, let me share. So this is what we get once that, that session is done.

   29. AGAakash Gupta31:17

       Wow.

9. 31:17 – 33:35

   Why dashboards aren’t enough: moving from insights to actions

   1. HMHarish Mukhami31:17

      This is the dashboard. This is the result of p-part one. This is the customer success AI agent dashboard only. Uh, we have the four tabs that we asked for. So customer analytics, it talks about your overall, uh, customer view, uh, how they're doing re-- from a revenue perspective over time, customer segments, top ten customers. Uh, click on customer funnel. Uh, it's more fancy charts around conversion rates through the funnel. Um, engagement and retention, uh, user engagement me-metrics, NPS, active users, retention rate, average users, uh, feature adoption. Again, all of this is coming from, from test data. You can obviously have your live data connected to it. Uh, last piece is the churn risk. So we do have a, a table that defines, um, what is the churn risk, how many customers are at high churn risk. Now, all these, uh, we're pretty familiar with. What is the next step in terms of looking at this dashboard if you're a customer success agent? Now you have to interpret this data and then understand, analyze, gain insights, and then take actions. That's the key part of this. So for that, the, that's the onus is on customer success agent to do that. So, um, in this case, what we wanna do is we wanna... This is, um, we wanna write an agent that is going to actually do that for you. So at the end of it, if you come back to the dashboard, there are actions available for you that you can take as a human in the loop. So this is the conclusion of part one of dashboard, and then we can go on to building, um, actually the first set of agents, which are going to analyze the data and, and then, um, take actions for you.

   2. AGAakash Gupta33:10

      So what we've built so far, it has built all of these charts via code. So it's not using a BI tool or anything like that, and it's created as a live website that you could theoretically hook up into your database so that you would already have these dashboards. So that's start one-- step one, the insights that a customer success agent might develop. Now we're gonna take action on those?

   3. HMHarish Mukhami33:31

      Yep, absolutely.

   4. AGAakash Gupta33:33

      All right.

10. 33:35 – 37:19

    Adding an ‘agent actions’ table via Cursor → Gibson MCP

    1. HMHarish Mukhami33:35

       I mean, the barrier to build, uh, is so, so low that I feel like h-how did we not have this [chuckles] two, three years ago? I wish I had this few years ago, so. Uh, let's move on, uh, to building the, um, AI agents that analyze the data and recommend actions. So, uh, in this case, a human either approves or takes actions based on agent recommendations. This is, like, the human in the loop use case. So, uh, first thing we-we're gonna do is we're gonna go back to Cursor and ask, um, Gibson via MC-MCP to add an agent action table to the schema. So what we did was we built this dashboard. Now we need to store those agents. So whatever the agents are doing, we need to store that. So we're gonna ask, um, Cursor to do that. And this is where you have the MCP integration with Cursor, uh, using Gibson, and then, uh, we're gonna do that right from the Cursor workflows. So let me share my Cursor screen here. So here, what I'm gonna tell the agent to do here, create a table in this database, uh, called agent actions with the following fields: customer account name, contact email, contact person, industry, churn risk, churn reason, action, message. Um, so essentially this is where the agent can write those actions for users to take a-and approve, uh, from the dashboard. So when we started building in Gibson, we used a web interface. Now we're actually building tables right from Cursor ID a-and deploying the database.

    2. AGAakash Gupta35:09

       So you can interact with Gibson either way?

    3. HMHarish Mukhami35:12

       Absolutely. Yeah. Yeah.

    4. AGAakash Gupta35:14

       And what would Gibson be replacing here if we weren't using Gibson?

    5. HMHarish Mukhami35:18

       You would be using, um, your traditional database, whether it is like the Postgres, uh, MySQL, um, SQL Server, if it's relational, non-relational MongoDB. Um, but the onus is on you to, uh, take your data model, deploy it in the database, manage the database infrastructure, ensure that it's scaling, it's optimized. Any sort of changes, you gotta do migration, um, and Gibson can do all of that automatically for you.

    6. AGAakash Gupta35:48

       Got it.

    7. HMHarish Mukhami35:49

       So it's getting schema, and then it's gonna go... Oh, it's actually adding a function to the test data though to populate, so instead of, instead of actually creating the table. So I can ask it to make, make sure do this. Uh, don't do this via Python script. Uh, don't create a Python script. Oh, it's actually doing the population. So, um, let me say, "Did you deploy-- redeploy the table?"Again, it's, it's checking the status with Gibson AI, and if it is deployed, if the radius schema is deployed, it's gonna redeploy it.

    8. AGAakash Gupta36:49

       Got it.

    9. HMHarish Mukhami36:50

       Okay, it looks like it redeployed it. Awesome. So one of the-- one quick thing I'm gonna do is go to Gibson, Gibson and show you quickly that agent actions table. We just asked it in Cursor. I have created a new agent action table here in Gibson. So let's see where it is. Which is right here. So we just asked through Cursor, it's available, redeploy.

    10. AGAakash Gupta37:18

        Got it.

11. 37:19 – 40:27

    CrewAI setup: multi-agent pipeline for churn detection and mitigation

    1. HMHarish Mukhami37:19

       So now we're gonna... What we're gonna do is we should a- we're actually gonna go back to Cursor, but we're gonna use the CrewAI framework. So that's the framework that we're using. CrewAI is, um, AI agent framework, um, like LangChain, if folks are familiar with LangChain. So, um, let me... I have pre-written this AI agent because it takes a, it takes a bit to do it. It takes few hours. It, it doesn't take days or, um, days or months to do this. Um, so what... I'll talk about how the overall setup is. Uh, there, there are two conce- concepts here. One is agents. These are the agents that you're writing. Um, if you look at, uh, this file, there are four agents. So first it's a database query agent. So the job of this agent is to query the database, um, the database that we have, which is a CSADB one, and then extract data and analyze information here. So, uh, your expertise includes identifying high-risk customers through, let's see. What does it say? Uh, running sophisticated, uh, high-risk customers. Where'd it go? High-risk customers through multiple indicators that might signal churn risk. So low activity levels, declining usage patterns over time, numerous un-unresolved support tickets, uh, low satisfaction scores, decreased purchase fre-frequency, long periods of inactivity, failed payments or billing issues. So you have the opportunity to actually pinpoint exactly what you're looking for. And then once it's done, um, it, it surfaces the data to the churn analysis agent that's going to analyze the churn, and then a mitigation agent that is going to provide mitigation stra-strategies, and finally, a data ingestion agent that is going to take these actions and put it back in the table that we just created. So there are four agents that are doing the job of what a customer success agent would do typically. Um, and then agents have tasks. So agents call the tasks. So these go into detail around what exactly you're supposed to do. Uh, and then all of these are put together as a crew of agents to finish the task. Um, there's a bit of a setup involved here, so which I have done this. So what I'm gonna do is I'm gonna just run this agent now. So, um, actually, before that, one thing that I wanna show you is now we pointed this to our database, right? So and then how does a Crew agent knows the database context? If you have a different data model, how is it gonna work? So what we have done is, um, it's pretty dynamic. So you can go, uh, look at... Let me see where the Schema tab is. So there's a, there's a thing called Schema here. Uh, this is a schema that you can copy/paste off your analytics database. So Crew knows the context of it. So what I can do here is, like, can you update the schema with the schema of CSADB one, use the Gibson AI MCP server. So if you're u- if you want this to run on a different database, you just say that, and then it's gonna go update that for you.

    2. AGAakash Gupta40:26

       Okay.

12. 40:27 – 44:31

    Running the agents and populating recommendations in the dashboard

    1. HMHarish Mukhami40:27

       And then I'm gonna kick this off. So it's gonna retrieve the project. It's, um, it's interacting with Gibson via MCP tool, getting the schema definition and automatically updating it. So you don't even have to copy/paste that here. That's done. So what I'm gonna do now is run the main, um, agent file here. So, uh, I'm just gonna hit, kick this off, and it's gonna go execute the set of agents to do the job that, uh, we have indicated here. So, uh, now the agent is in motion. It takes about three to five minutes to execute. While it's executing, one of the things I wanna show you is... So if you look at the customer success actions, I click on Agent actions, this is empty. Um, there are no actions here. Um, the customer success agent that is running is going to populate that. So let's go back to the Crew. It's, it's pretty interesting to watch and follow that run, um, in sequence and, and understand how the agent is actually working. So now it is kicking off the churn reasons agent, so churn mitigation agent.

    2. AGAakash Gupta41:46

       Which is one of the most important things our customer success agents are doing.

    3. HMHarish Mukhami41:50

       Absolutely, yeah. Um, one of the cool thing is obviously these agents are running all the time. So these are agents that are running twenty-four/seven. Uh, they never take day off or time off. They're always on top of it, very consistent, so. Um, one of the challenges that we face is even though you provide dashboards, and those are all handy, I think the onus is on the individual to actually ma- take those actions and be on top of things. Um, some are really good, some are not. So, uh, we typically don't have a problem with, with agents running this way, especially if we have strong evals backing those as well.

    4. AGAakash Gupta42:31

       Yep. And have we created those evals or is that a next step?

    5. HMHarish Mukhami42:35

       No, we ha- we don't... That's not part of this demo, but we should definitely do that. As a, as a... If you're running AI agents, you should definitely have evals to make sure that your is-- your agent is doing as it's intended, and then any sort of feedback that you want to give the agent.It, it's just like your, um, traditional end-to-end test cases, uh, that you would have for your typical SaaS offer as well. So-

    6. AGAakash Gupta43:00

       Yeah

    7. HMHarish Mukhami43:00

       ... it's done. So let's go back to, um, the dashboard. So it populated a bunch of actions here, so let's look at these. For this customer, specific customer, the risk level is nine. The action Crew is recommending... Actually, let me click on that. So action it is recommending is the potential churn reason could be product usage difficulties. Send email to customer, schedule training session, provide product usage guidance. So that's the first one. So let's pick something else. Okay. Second one is product team for escalation. So, um, unaddressed customer issues, send email to customer, and send email to product team for escalation. So here, it's recommending the CS agent, like, "Hey, send an email to customers," um, as well as like, "Hey, g- reach out to product team and make sure that we fix those issues." Uh, and similar, like schedule train... So some, some of these as like send email to customer, schedule training session, provide product usage guidance. So all of these, it has even composed the emails for you. All you have to do here is you can either copy/paste and then just send that email to that specific, uh, customer or product team, or you can connect your tools to take actions right off from here. That is a step two where the agents are taking, uh, agents are providing recommendations for actions here. Um, let me see. Now, um, the last step is, uh, is it's the part three. So

13. 44:31 – 53:20

    Going autonomous: adding email sending + Jira ticket creation tools

    1. HMHarish Mukhami44:31

       we evolved from building dashboard in part one, where humans use the dashboard as a tool to analyze data, uh, to building an agent that does the data analysis and recommends actions. Now for part three, we'll go a step further and add two more agents to execute those actions. So we have seen sending emails and re- sending emails to product teams, so we'll do one better and say like, send emails and actually go create Jira tickets. So why send emails? Like put a ticket in the, in the, in the Jira board. So-

    2. AGAakash Gupta45:01

       Okay

    3. HMHarish Mukhami45:02

       ... um, we'll do that. Uh, for that, what we're gonna do is we're gonna go back to the Crew agent and add those two new agents. Uh, let me... So here's where the agent just ran. What I'm gonna ask it to do here is, uh, now for the case where it's product issue is the churn reason, create a ticket in Jira. So write a Crew, uh, Crew AI Jira tool. This will be a new task, um, and agent combo, uh, and a tool. Uh, use the Crew AI documentation. I like to specify documentation, especially if it is like an up-and-coming tool, so that it, we get more con- it, we get more context. Uh, to Cursor, please add an email agent as a final agent without changing the existing agent tools. Email agent should send emails based on the churn mitigation agent results. Um, write a Crew AI email tool. Use SendGrid for emails. Otherwise it's gonna spin up an SMTP server, and you don't wanna do that. And then finally, like again, documentation is here. And then I will add the context of the entire project here. So I'm gonna give it the whole folder, and then I'll just let it run. There, as you can imagine, there's a bit of a setup involved, so you ha- you need to have SendGrid account. Then you have to go copy/paste that into your ENV, and then also set up your, um, Jira board, uh, and then also provide, create an app in Jira and then provide that app, uh, ID as well as, I believe, API token, uh, for agent to actually create those. So it's gonna go write those agents. In the meantime, what I wanna show you is, um, is my email and the Jira board so that we can see what the state is right now and what it will be once we write the agent and run it.

    4. AGAakash Gupta47:01

       All right.

    5. HMHarish Mukhami47:02

       So this is my email. Um, I have... W- while we are using customers' emails from, from the database, um, I wanna default it to using my email so that we're not sending emails to random people. So, uh, we're just gonna let use my email as default. And then let me see. And then here's my Jira board. So I created a CSA demo. There are no, uh, tickets created right now. And then now I'm gonna go, go to the, go to Cursor and show you the completely done customer success agent, which is autonomous. So let's go look at the agents here. Um, these are the ones that we're familiar with, like database query agent, churn analysis agent, churn mitigation agent, data ingestion agent. So these are the four agents that we w- that we saw and we ran. Now we have the email delivery agent, which is even communication specialists send personalized emails to high-risk customers based on the mitigation strategies. So there's a backstory to it, and then you have a bunch of things that, um, you wanna ensure that, uh, agent does. Preparing emails for batch sending, uh, knowing when to use test mode versus production sending, ensuring compliance with email communication best practices. A bunch of things. Uh, and then the Jira ticket agent. So, um, role is a product issue manager. Uh, create Jira tickets for customers experiencing product-related issues. So essentially this is gonna create a summarized Jira ticket for you.

    6. AGAakash Gupta48:39

       Okay.

    7. HMHarish Mukhami48:42

       Um, in this case, what I'm going to do here, uh, I think we may have to set up the environment. Let me try this. Uh, it would beSo here, here is where we're gonna be installing virtual environment. And then activate it, and then we'll do install the dependencies. Oh, I need to go to the right folder. It's gonna install the dependencies, and then can do Python main.py, which is gonna be-- which is gonna kick off the agent here. This again will take, um, a couple minutes, uh, to run. At the end of it, we should see Jira tickets created in the Jira board, and then we should see the email sent out on behalf of the agent. So this completes the loop around like, hey, automatically analyzing, identifying actions, and actually taking those actions. Um, you could imagine, um, we can p- very easily add a lot more agents that are very specific to your workflow. Um, if you think that, um, agents are finding your product expensive, and if the reason is like, "Hey, we found some other place that is cheaper," there's an opportunity for you to send some discounts and coupons if you have a, if a, if you have a bunch of, uh, people in there. Uh, if there are concerns around, around competition, you can definitely send the material around, like, this is why we're better than others if people are looking to switch to others. Um, and the cool thing about this, we're not doing this after people have churned and trying to identify things and trying to get them back. We're doing this much before when they're churning, so that is the key part here. And it's always hard for CS agents to be proactive because there are so many things that they could look at, um, so running a customer success agent like this is going to catch customers before they churn.

    8. AGAakash Gupta50:51

       Yeah, and I see this as the way we've created it, especially with the human loop check on some of these things, as a way to make your CS agents that you have now, like, 10X more effective.

    9. HMHarish Mukhami51:02

       Absolutely. I feel like customer success is, is a function that you do need human-in-the-loop. That's why I broke it down into a dashboard and an agent, uh, with a human-in-the-loop and then autonomous, so that you're not sending some random emails and that is only gonna make problem worse. So we wanna make sure that, uh, somebody's actually taking a look at these.

    10. AGAakash Gupta51:24

        Yep. But you could theoretically, if you wanted, like we were talking about, write some evals where an LLM is evaluating what this one is doing and okay it too, if you wanted to go the other step.

    11. HMHarish Mukhami51:35

        Yep, absolutely.

    12. AGAakash Gupta51:36

        And you could just use Cursor like we have been with seems like text language prompts to get there.

    13. HMHarish Mukhami51:42

        Yeah. Nowa- nowadays, um, almost pretty much a lot of tools have MCP servers, so you can definitely connect them and then execute your workflow. Not just Cursor, any MCP client, any agent with an MCP client that can support that would do it.

    14. AGAakash Gupta52:01

        And we talked a little bit about CrewAI. What was the role here of that tool?

    15. HMHarish Mukhami52:07

        So CrewAI is the AI agent framework that abstract away a lot of things, um, that you would have to do from a memory management, from a tool definition perspective, from chaining of those, debugging. Um, all of that is, is done by Crew. It's, it's, um... I feel like it's a, it's a more, uh, more, uh, trai-- I, I guess more relevant for maybe semi-technical folks as well because it's fairly straightforward in terms of how it is set up. And once you know it, you can tweak the backstory, you can tweak what you wanna write from a prompting perspective on the agent's, uh, side to get the Crew to do what you wanna do. There's also other tool, other frameworks called LangChain. LangChain I think is a bit deeper, um, but I felt like this is more targeted towards maybe semi-technical folks. Um, definitely technical too. So this is done. Um, so the execution is done. So, um, it took maybe, like, three minutes, so you, you're free to choose whatever framework you want for this demo. I felt that this, this made a lot of sense.

    16. AGAakash Gupta53:12

        Mm-hmm.

    17. HMHarish Mukhami53:12

        Especially since I have a PM background as well, so, uh, while I'm technical, but I'm not, like, a pure developer.

14. 53:20 – 1:01:46

    Results: Jira issues created and emails sent; broader ‘AI employees’ vision

    1. AGAakash Gupta53:20

       Yep. So let's go check those Jira boards and email.

    2. HMHarish Mukhami53:24

       Yeah, absolutely. So this is the Jira board. It created three tickets, one, two, three. Um, let's click on this. Customer is facing unaddressed customer issues with the product usage is currently unresolved. Eight tickets is critical to their continued engagement with the product. If we had very specific ticket information, we could have gotten this, but this is essentially, like, created, um, as a bug. I think one of the things that was cool about this is it put in labels saying it's churn risk customer-reported, so you can obviously make a very significant prioritization decision here. So, uh, as a product team, we just created this. Um, emails can sometimes take a bit, but actually they showed up too, so, um, uh, four emails. Let's look at this. Um-

    3. AGAakash Gupta54:14

       Look at four.

    4. HMHarish Mukhami54:16

       Uh, "I hope you're well. I want to sincerely apologize for the unresolved issue. I would like to offer you a discount." Wow, okay. [laughs] This is, uh, uh, went a step further. Obviously, we didn't put in, um, specific customer success agents, and we could easily configure to have this coming from them versus me. Uh, and then, "Let's resolve issues together, improving your experience with our product. We're here to assist you." So, um, it sent out four emails, and then at the same time created three Jira tickets here, uh, on the Jira board.

    5. AGAakash Gupta54:46

       Amazing. And that covers our live cooking session. So today we created an AI customer success agent. Using the concepts we went over today, what other AI employees could people create?

    6. HMHarish Mukhami55:02

       Um, I think-I think pretty much all of the information work is, is up for automation. So, um, I think you covered product manager before. SDRs are a very popular use case. Um, financial analyst-

    7. AGAakash Gupta55:18

       Sales development representative.

    8. HMHarish Mukhami55:20

       Exactly. Uh, sorry, say that again.

    9. AGAakash Gupta55:23

       SDR, sales development representative.

    10. HMHarish Mukhami55:24

        Yeah, yeah.

    11. AGAakash Gupta55:25

        The person who goes out and gets the initial lead. [laughs]

    12. HMHarish Mukhami55:29

        Absolutely. Um, there is also meeting preparation agent, uh, executive assistant. Uh, you can chain together all of these agents to, to create a specific role that is tailored for you. So, um, we also have, um, AI agents that are, uh, acting as recruiters, so they do a lot of upfront work around analyzing resumes and, and picking those based on the JDs. So, uh, these are all very popular use cases. Anything that involves putting together, whether it is structured or unstructured data to take action, like they're all up for, up for automation.

    13. AGAakash Gupta56:08

        Wow. So basically, the criteria here is you can build an AI employee for any task in your business that involves ingesting information and then sending out information. So all of those tasks, those human connector tasks, we saw it create Jira tickets, we saw it send emails, we can now do. That was an awesome demo. I wanna spend just a few minutes here at the end learning a little bit more about you. So you worked head of product at Siri, you were CPO at LeafLink. Talk to us a little bit about how you came into this process of creating Gibson.

    14. HMHarish Mukhami56:47

        Yeah, absolutely. Um, so before Apple, I spent almost a decade at Microsoft, and, um, and one of the core focus for me was building dev tools and developer frameworks. Uh, primarily worked as a developer at Mi- on Microsoft Dynamics, and then was a PM in, uh, on the same product. So, um, I saw firsthand, uh, some of the biggest challenges, uh, especially around, um, building software, um, building infrastructure around databases, and a lot of those are still not solved. And then those problems continued, uh, on to, uh, my next roles at, at Apple or at LeafLink. So the biggest impediment to delivering software quickly, uh, solving customers' problems quickly is delivering software, and that is relying on engineering teams to develop that. While we ha- we are not at a place we can say that engineers are, are... We don't need engineers and AI can write code for you. I don't think we're there yet. But we have enough tools to make engineers, like a, a junior engineer, a senior engineer and, and, and in general, any person, like you, you said before, a 10X version of themselves. So, um, that's the impetus behind, uh, building Gibson. Uh, we are focusing on the database layer of, of the dev stack. So we're focusing on all the way from, uh, design, uh, development, deployment, management, and scaling of databases, which coding tools don't do today. So, uh, maybe coding tools can write data models and schemas, but there's a lot more involved in database management than just writing, uh, schemas and data models. So there's a big c- component around, uh, schema migration, data migration, query optimization, um, impact of any schema changes you do, and we're, we're handling all of that via Gibson. Uh, we currently solve, uh, from a PLG perspective, zero to one use case. For enterprise customers, we can, we can solve for existing databases as well.

    15. AGAakash Gupta58:49

        What's your grand vision? If everything works out for Gibson over the next three to four years, what will people be building on Gibson?

    16. HMHarish Mukhami58:57

        Yeah. Uh, my vision is that every single, uh, business, individual, uh, building technology products, AI agents, data-powered pro- uh, products are using Gibson as a backend, um, to manage, um, databases, scale infrastructure, and, and be on their journey from going from zero to one to one to hundred. So whether you are serving 10 users to 100 users to tens of millions of users, Gibson is there with you on your journey, and it makes it very easy for you to scale. Uh, one of my favorite use cases with Gibson, Aakash, that you have just seen, is you plug Gibson into vibe coding tools, um, um, and you're not building a prototype anymore. So you are backed by a very significant production-grade database. So if you-- It's already deployed in the cloud, so you publish this and you get 10,000 users tomorrow, we can handle that. So you don't have to scrap this and rewrite it. So, and it's very quick, right? We provisioned database and deployed it in a matter of seconds. So these are the things that take a lot of work. And one of the, one of the key things also is it's not enjoyable. Not every developer enjoys managing infrastructure and databases. They just wanna write code. So we handle the grunt work.

    17. AGAakash Gupta1:00:15

        So is it fair to say that with the tools that we've showed you today, you can go from beyond just prototyping to vibe coding with a production-grade database to actually put something out there that acts as an employee for your company?

    18. HMHarish Mukhami1:00:29

        Absolutely. Um, the one thing I would say, whether you're vibe coding or whether you are, like, um, using database as a copilot, uh, we haven't done that today, uh, i- in the interest of time, but any code that AI generates, please read. Uh, make sure that you don't just blindly accept. Like, make sure that it is the code that you want. I typically don't, um, code via in the agentic mode. I know a lot of developers on my team, they definitely use AI, but they use it in a copilot mode to complete the sentences versus, like, actually writing code for them. So, uh, make sure you're, you're reading what you're, uh, what you're actually accepting.

    19. AGAakash Gupta1:01:09

        All right. Harish, if we had another hour, we could do another hour. Thank you so much for helping us build an AI employee today.

    20. HMHarish Mukhami1:01:18

        Thank you for having me, Aakash. It's been a pleasure.

    21. AGAakash Gupta1:01:20

        I really hope you guys enjoyed that episode. It would mean a ton to me and the team if you could please subscribe on YouTube, follow on Apple and Spotify podcasts, and leave a rating and review. Those ratings and reviews really help grow the show and help other people discover the show, and they help fund the production so that we can do bigger and better productions. Can't wait to share the next episode with you. Until then, see you later

Episode duration: 1:02:03