How to build AI agents to automate your work

00:00
sully
A website in five minutes and I'm going to try to do this live. I haven't set anything up and I'll walk you guys through how I would set it up, how I would run it. And I don't know, one percentage of the audience is like code doesn't code. So I'll show you this one. If you were to do it's going to have a little bit more involvement. But a couple of things before I dive in, maybe in the audience, if anyone has any basic apps that they want to try. In the meantime, I'll just go over the tools that I'll be covering. So for those of you who maybe aren't too familiar with the AI codering tools, there's a bunch for me. I use Vercel v zero and actually Claude 3.5.

00:46
sully
They're probably the best available AI tools to help you code whatever you want. And I'm going to go into explaining what Vzero is in a second. But the reason this workflow works is actually both Claude and V Zero are trained on very popular libraries. The nice part is because these models are all trained on that. You can quite literally ask it to write something for you, copy it and paste it into your app without any configuration or anything like that. So let's see here. Okay, cool. Got a lot of people. Yes. All right, so let's go into a live demo, and I hope to God that this goes well. I'm going to try to. So I have this. Okay, so I have this demo workspace here. We're going to launch it. It's just like there's nothing here, actually.

01:50
sully
And then I'm going to go to cloud AI so you guys can see how I do this. I'm going to v zero. So these are the two tools that I'm going to be using. And this is a live demo. So we'll time limit it to about five minutes and we'll go into launching this workspace. Cool. So for those of you guys, if you guys use GitHub, you can just launch a coding workspace where you don't need to set up. It's similar to relet. I don't know how many of you guys use relet. You guys could use replay too, if you don't even need anything. This is all in the browser. This is basically, I go into here and would just say, hey, help me set up a blog page using Shagsian.

02:52
sully
So actually here it's going to go ahead and generate me the thing the other way. I like to use a lot of these AI tools for building sites when I'm coding is actually coming to these and using this as an initial ideation here. I'm going to say a homepage for a blog site. So let's go here. Installation next. Js. Go copy, paste it into here. Bigger terminal settings. All right. Allow. Yes. Nice. Install. Install. Got some nice designs here. Cool. This looks pretty good. So I got my initial set up here. Again, all I'm doing is I'm literally just copy pasting all of this. We'll go copy this into here. Let's see here. Yes, the demo gods have not decided to be against me today.

04:22
Ben Tossell
Yeah, it happens. And then obviously because you're a coder, you know exactly what things you are copying and pasting right into certain places.

04:30
sully
So one of the things, yeah, actually that's a good point. So maybe if you're maybe unfamiliar with like kind of what you're doing, and one of the ways I like to do it is say, you know, really just push Claude into being my personal AI coding assistant. So like one of the things that I would straight up do is like, I will straight up screenshot some stuff. So I'll come here, I'll just screenshot my page and then screenshot, I'll copy this, come to Claude and I'll paste it in. I'll say, hey, help me set this up in my terminal. In my whatever terminal. I don't know how to code. I guess it has a nice little screenshot of there in the background.

05:16
sully
And what it'll do is it'll actually give me the exact directives and explanations for how I would get started in something like this. So it says here, let's get started with a new project. Copy this into the terminal. I always have ran this once, but let me try running it again. Oh, actually, I know what it is. See my app then here, actually it gives me the directions. This is where I messed up, where it says, all right, you're going to make this app. It's going to be called my blog. And then you're going to run this command and then I'm going to copy here. So I'm not even going to read the directions from the site. I'm just going to listen to what Claude tells me. Yes. And then it says here, I can pick the styles. Great.

06:08
sully
So we'll let that run while this installs great. Now it's there. Now I can do okay, if I say, how do I add components? And Claude will just explain to me exactly every single terminal command that you need to know. And I don't even need to read the documentation here, which kind of explains what I would normally do. So I can go like this and I can install it and tells me how I can add the different components, how I can run this, and I'll try to wrap this up and move on to the next example. But if there's any questions, happy to answer them, because I have a bunch of demos and live examples. I don't want to spend too much time on anyone. So.

07:11
Ben Tossell
Yeah, sounds good. Yeah, I think we can move on to the next one and we'll collect questions for the end. So if you have questions that you want on a specific demo, yes, just have them ready and then we'll do it.

07:31
sully
I'll leave this here while this is all doing, and we can come back to it. But yeah, that was just an example of how I myself, as someone who does code, but you don't necessarily need to, you can come in here and just literally take screenshots of what you're trying to do and just give it to Claude and it'll sit there and just guide you. You don't even need to know how to code. And that's what's really so exciting about using these kind of tools, is that normally this would have taken you, like, for me, like a year ago, probably like a day or two, just to set everything up. Now I can do it in 510 minutes with Claude. Cool. Here we did the coding example, unfortunately, did not do it in five minutes. Maybe I was being a little too ambitious there.

08:10
sully
Let's move on to another example. This one is a little bit simpler. It's going over how we can maximize, whether it's Claude chat, GPT, or Gemini. The way that I like to use these systems to give better outputs by multi step prompts and change, so that they can iteratively improve it. The reason is that when you give models the ability to think and you give them a little bit more time to execute a task for you, they actually do a lot better. So in this demo, I'm going to be covering cloud models, and you can use this for chat GPT as well. But I've just been using cloud entirely. Stop using chat GPT. What you can do is you can use tags. I'll show you exactly a demo of that.

09:03
sully
The way that you do this is that it lets the model think and it lets it figure out what it can do. I don't know how many of you are familiar with this whole think step by step. I'm sure you guys remember with tipping. Please sir, tip me asking the model to tip and stuff like that. But really the short explanation is if you give the model the ability to explain its thoughts and its reasoning, it performs better. I'll go over how I do it for a couple of examples and how I set up the prompt. And again, if you guys have any questions, we'll leave it for the Q and A of like hey, I wanted to do it for maybe transcript summaries or writing emails or anything like that, but I'll go over one example and then show you exactly what happens.

09:46
sully
We'll go here, we'll start a new chat. Okay, cool. We'll say, maybe I am using this to write an email. Again, super simple example. And this is also part of partly a lesson of how I write prompts and how I get these systems to listen to me, but also how I would use XML tags. Let's say I'm a busy professional writing emails that hopefully don't sound like robots. It should be personalized and sound genuine. Sure, this is just some pre prompting, but now this is kind of what I was referring to about using XML tech. So what you usually tell the model is, before you answer, please use a specific structure so that when it's responding, it will think through this structure and then give you the output. So let me give you an example.

10:46
sully
When you, when you reply, always use XML tags to show your thinking. Start by including your thoughts and reasoning, like so, thoughts. And then I'm going to say also. And then reasoning. And the actual email. So this is sort of the format that a lot of us have set up is you say, all right, give me your thoughts, give me your reasoning and then the actual output. So then when I ask Claude to do this, instead of just praying to God that it's going to get me the right email, what it'll do, it'll kind of go through the thinking chain and it's really useful because you could kind of add more prompts. So, for example, if you wanted to say, I want you to think in a specific way, you'd say, well, I want you to think about this specific thing.

12:04
sully
Or maybe I want you to think about your tonality and then I want you to give your reasoning for your tonality for why you know, you're replying. And then write the email. So maybe I say, you know, write a email to Ben saying how much I enjoyed his AI. What's this called? It's the, so what is it called? The session is AI session.

12:38
Ben Tossell
Yeah, workshop session, whatever.

12:40
sully
AI workshop. Workshop. All right. And just to show you how you would do this, obviously this is kind of random, but as you can see here, what it does is it gives me its thoughts. It says to write HTML, I need to kind of give me any additional prospects, any thoughts, and then it gives me its reasoning. So it says here, given the request for a non robotic, personalized email, I'll use a casual yet professional opening. Mention the specific elements in the workshop, et cetera, enclosed with a different one. And this is one of the prompting techniques that I like to use in my workflow for maybe more complex tasks.

13:21
sully
An example is if you're maybe asking it to write a blog for you, or you have a giant call transcript, if you just ask it to, hey, what are the top insights of the call transcript or whatever, you'll notice that a lot of models, they won't actually give you maybe those golden nuggets that you're looking for. By setting up this xml, this sort of, these sort of tags and thinking process, what you can do is you can force it to go in a direction to extract those insights. So you could even add in that example, you could be like thoughts, reasoning, key insights, key extracts, and then finally your output. And it'll kind of go through this chain and it'll actually significantly improve kind of what you can get out and maximize the output of these sorts of models. Yes.

14:07
Ben Tossell
This kind of shows me that it's one a way to look at what a model does when you give it a prompt. So then you can go back and rejig what you actually say. And say, actually, I didn't want that one weird thing expressed in 2000, gratitude, for example. So it gives you like some feedback in how you can then go and prompt better. But also then, like you said, if you have thoughts and you know that you want a summary to go a specific way, you guide it that way, right?

14:36
sully
Exactly. And when you can see what the model thinks, it's a lot easier to guide it. So here, for example, like you said, I, maybe I don't want it to express enthusiasm or gratitude. I say like, no, I want you to be cold and demeaning, whatever. So, and normally if you were to sit there with the model, right, and you would have to like really try to like wrestle with it and be like why is it not working? But this way you're able to kind of see its thinking process and easily guide it and get better outputs in less time. All right, so another one I have is something that I've been using a lot of which is using video.

15:14
sully
So I don't know how many of you use Gemini or how many use video, but it's something that I've recently started to do a little bit for more of the ideation, brainstorming. And the reason is video is just way easier. I don't need to sit through and type out what I'm trying to say. I can just record a video of myself and what I'm trying to do. So I'm going to try to do a live demo again. This is just all live and showing you maybe what I would use it for. Usually Google Gemini is pretty good, but let's give it a shot. So here I have it. I'm going to record my screen. Hopefully it doesn't mess with anything zoom related.

15:57
sully
In this example, what I'm going to do is walk you through, I'll go to some landing page, I'll go to my company landing page, I'll walk through how I would walk through an AI. And then essentially what I do is explain the product and then I'm going to ask it for a critique. So I'm going to say, hey, this is my product. This is what we do. This is my landing page. Can you please help me improve it? Maybe ask for a specific marketing copy that I can improve to improve my blank. So traditionally if you have a website, a lot of people's workflows is like you'll copy the content, you'll paste it in half the time, the AI won't know what you're talking about because it has no clue what your site is.

16:34
sully
Maybe you'll paste in the link and it'll grab the content, but it really doesn't actually understand what is on your site. So let's try running it through an example and I will screen record and let's see. Okay, I will record my screen. I'm going to minimize. All right, so this is super meta of me on a call recording myself, but we'll see how it goes. All right, so let's actually go here. I'm going to open a new tab so I'm not signed in. All right. Hey, Gemini, just going to be going over today. My company, so my company is called Auto. We basically took AI agents, we put them into tables. What I need help from you. Today is critiquing my landing page and explaining to me what are some areas that I can improve on it.

17:27
sully
So I'm going to walk you through the product and go over the landing page. I'm going to ask you if there's any copy or anything like that you can improve for me. So this is autogrid. We basically built a tool that helps people actually work with AI. We didn't like chatbots, so we decided to go for a table approach. You basically can automate hundreds of tasks in minutes. I like this initial part of the landing page, but I'm not sure if maybe having explore templates right underneath is good. So curious what you think. So here I have a little whatever you would call this, I guess the tabs to maybe go over use cases, although I think this could be improved a little bit. And then here I have the features which go over it.

18:12
sully
You can go ahead and read all these features, let me know what you think. If there's any areas that I could improve on my landing page. All right, so we'll save that. I'm going to go back to AI studio and just give me 1 second here. Let me grab that video and upload it to Gemini. Awesome. So then I will just go here, upload video. Awesome. And actually, you know what we'll do is I'll actually go and show you that technique I talked about the tags and I'll actually ask Gemini to do that technique to show you how you can get better output so you can give it a system prompt. I'm going to keep it minimal because it takes a little bit. I want you to critique. Oh, my spelling is just not it today. Critique my company and landing page.

19:43
sully
I want you to be harsh but honest. Reply using this format. Thoughts, thoughts. And then so here as an example, I'm going to show how you could basically say something along the lines of give me what I have. My thoughts are. So for example, let's say speaker thoughts and then maybe for now we'll just keep it simple. Critique, final answer and we'll hear answer. Okay, so we'll go here, we'll make sure I have the smartest model, Gemini slash, not the best one. I want Gemini pro and we'll go ahead and run it. So in that example, hopefully it doesn't mess up the first bits. Let's see here. So here, unfortunately Gemini doesn't do as well with kind of the XML tags here.

20:56
sully
But basically here gives me an honest feedback of me recording the whole landing page, scrolling down, giving very detailed context into what my thoughts were on the landing page and then finally uploading that to Gemini and saying, hey, please do make it better. Or maybe there's specific parts that I want to discuss with it. It's similar to akin having, I think, GPT 400, the vision, whatever demo they had with the Scarlet Johansson voice, it's very similar to that. This is a little bit more manual, but it's very useful again, because it makes that iteration cycle between you and the AI much quicker. If I were to sit here and talk with it and being like, oh, this is what I meant about our landing page.

21:42
sully
So here it gives me a bunch of errors and improvements and obviously watched the video and gave me an explanation of why it did it. And I can sit here and I can chat with it and it has a lot more context to my landing page or my site. And you can do this for your business. You could have a conversation going through what your business does, maybe your dashboard asking it about analytics, and it's really good instead of having to type it out and just add all that context. Cool. All right, we'll move on to the next one. So this was basically using existing flows, existing tools, nothing out there, nothing too crazy either, using Gemini Claude, and then obviously there's a v zero. And these are all, a lot of these that you saw today are not agentic tasks.

22:33
sully
These are the workflows that I use consistently, basically every day because they're reliable. I know that when I do this with the model, it's going to give me what I want. But then the next thing, obviously, I think a lot of people are excited about is agents. So let's go over what agents are and then we'll hop into some more demos. A while ago, I made this little image, and agents are, I think, the term agent. Everyone has a different definition. Some people think it's one thing, some people think it's another thing. But I like to think of it as a very simple thing. So a user, maybe you or myself, go to an LLM and I ask you the question then what it does is it uses a tool.

23:21
sully
Maybe I give it the ability to read my emails, which is underneath the hood, a function call. That response goes back to the AI. Then the AI says, all right, what do I need to do? Do I finish? Do I have sufficient information to answer the user's request? If I don't, I'm going to go back and continue this loop. I'm going to loop, but if I do have enough information, I'm actually able to finish and give the output to the user. In the simplest form, this is what I like to think of an agent, and now there's a lot of different variations. You have multi agent systems where they can collaborate on a topic. You have hierarchical agents with a supervisor and a manager, but in the simplest form factor, this is what I like to define as a single agent entity.

24:15
sully
How can you build agents? There's a lot of ways that you can manually build agents. You can use crew AI. Not sure how many of you use crew AI. You can use lang chain, you can actually use the Vercel AI SDK. They launched a couple of recent more. Again, these are all lower level implementations of agents where you have to code a bit. And then there's llama index. If you're someone who is very familiar, very technical, and you're like, hey, I want to dive in, I would actually recommend that you do go in and you try to build a couple of your agents in all three of these frameworks, because it lets you understand what they're good at and what you can use them for. But obviously if you're like, hey, I'm not technical, I don't really know how to code.

24:57
sully
You could use chat GPT, but I will say that it might be a waste of your time. You're better off using products that actually help you build agents. And actually you don't have to sit there and code them. All right, now to what we've been building at otto grid agents through tables. I have a bunch of demos for you guys, all live. But the simplest explanation is that essentially you come in and you define your table, which is a workflow directly within our product. Then in the product you can basically have a column, be your own AI agent, then you use natural language to build iterate, create columns, and then you can deploy hundreds of thems, hundreds of agents at scale. And what we focus on mostly is a couple of areas, mostly market research, scaling your outbound.

25:52
sully
If you're a team looking to scale your cold outreach, wanting to write more personalized emails and looking to enrich contacts so that you can send emails and other ones, we have a couple more examples of analyzing PDF's and stuff. So that's us. That's our main focus of building agents through tables. Let's go to some examples. I have a couple of examples here. I'll walk you through how you could set up scaling your personalized outreach, and how you can use otto grid and it's agentic platform to go out, research, look into topics, look into people and then actually generate personalized emails. Let's go here into autogrid. I'm just going to enable something for myself here. I have Apollo. Cool. This is the platform. This is autogrid. Very similar, I guess, ui ux initially, if you were to go to chat ship here, clock.

26:49
sully
But the way that we like to think of autogrid is that you have basically two things. You have your table and then you have the actual work being done by the AI. What we always say is that when you're coming into auto and you're using it to build a lot of these workflows, is making sure that your table is right. You iterate build and then you essentially, once you have your table, have the agents go to work. Let's go through an example of how you can do it for maybe like company or cold outreach or something like that. So I want to build a table to scale my outreach to companies. I need their, maybe their CEO company news, any recent funding analysis, and basically you can just describe what you want and go ahead.

27:51
sully
Here is where we have an interface where you can actually sit down and actually talk with another assistant to help build out your table. You can come in here, you can create your own columns. You can click here, I can create my own columns. We have a lot of different ways you can create a column. If you use, if any of you are familiar with airtable, it's similar style framing. You come in here. And the way that I like to think about it is, this column is a definition for an AI agent. What I mean by that is, I say this could be, let's look at the funding announcement here. What I'm saying is, this is the funding announcement. I actually have other tools. I want you to use the web search tool. I'm going to give you some inputs.

28:33
sully
The inputs, one is essentially when I'm giving context to this agent. So I'm saying use the company, which I will give you to go and find me their funding announcements. And then I give it a prompt. And internally, what you'll see once in a second is that we actually, every single cell becomes its own AI agents capable of thinking, reasoning and solving a very simple task. So, and here maybe I want to say, I also make a column for personalized emails. So here I can actually iterate with it and say a couple of columns. So do that and it'll operate in behind the scenes creating a column. So it says, okay, cool, maybe I don't need this web search tool. It says, all right, generate a personalized email, engaging tailored archives. Maybe I sell AI. Insert fill in the blank.

29:26
sully
What your company does here, research software. So click save. That's here and we'll go ahead and save this table. So then here I can actually, there's a couple of ways that I can get started. One of the ways is I can just manually add a list of companies. So we know we have a lot of people who have a list of leads or they have a list of from a spreadsheet or from HubSpot Salesforce. So you can come in here and you can paste it directly. There's also a way to import contacts, but if you're like hey, I don't have any list. We also have a web researching agent that will go out and go find your company.

30:03
sully
So maybe I'm trying to look for AI startups that recently launched and we'll go and do a little bit of web search, go find some AI startups for you. Right now it's limited just to five results, but you can have 30 results. I just don't have a pro account. Cool. Here, I just have them on this demo account. Found me a couple. We'll go ahead and click save and run. You see how something like this would work. What we do behind the scenes is I think earlier I mentioned about this agentic loops and multi step tasks where agents suck at. Right now, what we try to do to circumvent that is we like to break it down into these columns. Really, you could think of these as basically steps in a workflow, but in a table form.

30:53
sully
So as you saw here, I said, okay, I want you to do a personalized email. It actually highlights the columns. This one says, okay, I need the company, but this column will not run until all of the other columns are done. So that means that we basically have built in multi step reasoning directly into the table managed for you. So you don't need to do that. And the nice part is, if at any given moment any of these steps fail, I don't like the output, I can select here, I can rerun the cell, I can select all of these, I can run all these cells. We give you a very familiar interface to operate this. Then here for example, I have my cell here, maybe I'll just open this up a bit.

31:36
sully
It gives me my email, I can copy this, I can look at it here where I can, it uses all of that information from what it's gathered to help me generate a personalized email. Obviously you could come in here and configure the prompts, like what I mentioned before, where you can make it more personalized, but just to show you some of the things that you can do within autogrid. Again, the nice part is all here within a table. Soon we'll be adding the ability to just right click and send the email, but that's coming a little bit further down the line. I have a couple more examples, but I want to pause here because I know we have 15 minutes and I want to leave some time for questions. I have a bunch more demos.

32:16
sully
You can do market research, you can use web scraping. So the nice part is you can actually select types to be a link, and you can literally just give it a link and just say, hey, I want to know from this link the names of the attendees at the list, for example. Then maybe you have a bunch of links and you can paste those in and give it the definitions of the columns. The AI agent will go in onto the site, scrape the content, and put it into the spreadsheet for you where you can export into your, maybe your own spreadsheet, your own systems, things like that. It's a very familiar interface. Then the last one is, we do actually have, I'll just quickly go through just examples. We have a couple of templates here where you can actually analyze PDF's as well.

33:00
sully
I can go here, I can just quickly see this where I can upload a document for a sell, right? And then what I can do is I can say, all right, actually I want you to use that document and we'll run the reg, we'll analyze the PDF, extract it, and kind of grab all those relevant numbers for you. So yeah, I will pause there and leave about like 15 minutes or so for questions. Hopefully I didn't go too fast. I was trying to be wary of time and trying to get through all the slides.

33:27
Ben Tossell
So, no, that's awesome. I think the table I posted in the chat that for some reason seeing it in action in a table helped me understand it a bit more of like what the agent is doing. It's taking, it's like doing one thing first finding the company, then that context goes to the next agent to find the CEO. Then that altogether goes and finds all the other stuff. So just like, is there a best way to do that? I suppose. Like does it matter which way around you do it or not really.

34:03
sully
Not really. It's very intuitive in the sense that it's like, okay, in order to like, I'll give you an example of this one. So here's a bunch of call transcripts. Right. So the first thing I say is, I want you to go find the person's name from this call. Right. Who's the name of the person then? Actually, once you find the name of the person, I want you to go find their LinkedIn profile. So basically, it's kind of saying, okay, grab me the person's name, and then actually go grab me their LinkedIn profile. So it's being able to, from the call, I would say, like, you could ask it from the call to get me their LinkedIn profile. But breaking it down into these kind of steps, because it's useful for me.

34:37
sully
I'm looking at this like, I would find it useful to know who the person was, and then I can look at their features, requested, the summary, things like that. So it's very much like, again, what I was saying is that you kind of set up your workflow with like, one or two rows, figure out the form you like, and then be like, all right, here's like 100 call transcripts. Go and automate all of that instead of uploading 100 transcripts and kind of finagling it. And then like, you know, it's better to start slow. And then once you have your table built, let it go. Yeah, yeah, let me do all this.

35:08
Ben Tossell
Yeah. If folks want to start raising hands for questions, Gurpree just asked, is there a way to prevent it? Is there a way to prevent hallucinations with this, does this as this in a table format and does things one by one. Is that helping reduce hallucinations and pulling things from sources and everything else?

35:30
sully
Yes, actually, we spend a lot of time trying to tackle hallucinations. It's the number one thing that we look at. So obviously here, if you actually click on a cell, we give you a lot of different data points. We say, here's the reasoning for why we picked this total funding round, 130 million. We give you the sources so you can go in and verify. And we have a lot of internal mechanisms where essentially we will look at multiple sources. And if those sources, let's say three sources, corroborate on answer, we're pretty certain that's the right answer. But if literally none of them corroborate, we're pretty confident that's not the right answer. Or maybe we're not sure, and we'll actually give you the reasoning.

36:06
sully
So when you look at that data, you can say, okay, the agent wasn't really sure, and you can be a little bit more trustworthy. So it's something that we spend a lot of time tackling and we're still improving on, but our number one priority has always been minimized. Hallucinations. To get it to a point where you can basically look at the numbers and trust it, that's kind of our goal. We're not there yet, but trying to figure out ways that we can make the data more trustworthy.

36:28
Ben Tossell
Awesome. Makes sense. It's a question here from John. Oh, he's asking how you took the screen recording when you were.

36:39
sully
Oh, I'm on a Mac and I click the. If you're on a Mac, I just use the screen recording tool. There's just like one that I can just drag. You could use loom and you could record a loom and then download the video if you don't have a Mac. I don't know. On Windows, like usually windows is a little bit more annoying, but I'm sure they have a tool. But yeah, I just recorded my screen and downloaded it and uploaded it straight up to Gemini.

37:03
Ben Tossell
Yeah, that's what I thought as well. Yeah. Remember, anyone want to ask a question, raise your hand or if you're too shy, put it in the chat and I will ask on your behalf. Have you got an example table there of web scraping? Because I think a lot of people seem to be interested in that. Yes, that'd be good to see.

37:23
sully
Yeah. So let's actually go here. Let's hope I didn't delete it. I had a table here earlier in the day. Let me find it, maybe we can make one. So let's just say here, let's just make one. It's probably easier table to scrape a URL. Honestly I don't even need to, let's say here you can also make one from scratch. Sometimes I know exactly what I want. I just go here, make a table from scratch. I come in here, I say a URL to scrape, I give it a link type, I said the URL and then maybe I can just say, you can come in, you can actually define all the specific fields that you want.

38:09
sully
But for this game we'll just say summary of the page for now and we'll click that for the type, we'll actually go long text and then we'll click that summary of the web page. Cool, let's, so let me find like a page, let's just scrape this page. I go in here, I click a new row, I paste this in and it'll just run like that and it'll go scrape that page for me. Yeah. And then I can ask it other things as well. So maybe I say summary of the page. I can kind of start to create more columns, and I can say, all right, you know, maybe what are the key names on the page? And you can kind of start to create these, like, workflows.

38:58
sully
I like to think of it as like maybe some level of like, kind of programming as well is where like, you're doing it through the table, sort of like Excel formulas. It's like, okay, a plus b and stuff like that. So here gives me like, the content of this page. And you can do this with, like, I have a couple of examples, actually, now that you mentioned it with, like, for example, if you wanted to do like, zillow listings, this is like a, you know, a proper use case where here I have the listing. I can click on. This listing gives me whatever this is. This page gives me all this information, whatever. Then I have the price, the year built, number of bedrooms, monthly rents, bathrooms, and all of those are just directly defined through the table.

39:40
sully
If I just copy this over to mine, I can show you here. It's like, this is the price. And I describe what I want, the year the property was built, number of bedrooms. And I just said, use that link to kind of do whatever other things we have as well is you can kind of come in here, you could do classifier. You can have like multiselects. So in this example, like, I put like, you know, what's the interviewer's name? So this is either me or my co founder. And it can pick from the multiselect. The other one that's really nice is like, what's their use case? So I have actually a multiselect, and I can say, all right, based on the call transcript, what's their use case?

40:12
sully
And it'll kind of pick the best options and help me categorize and label my data as well. The agent will go and figure all that out.

40:19
Ben Tossell
Yeah, that makes sense. That's awesome. We've got a question from Alexander I've just asked to unmute, if you want to jump on.

40:28
Alexander Orlov
Yeah. First of all, thank you very much for these exciting demos. They're just fantastic. I have a quick question regarding your service, autogrid. Is it possible to somehow integrate it or maybe integrate it in Zapier or make so, for instance, to transfer data from the tables of otto grid into your automated workflows.

40:59
sully
So that's something we're looking at, how we can basically add automations built into the platform right now. The way that you can import and export data is just through CSV. So we're adding some more integrations. But I'd love to know what automations you have and how. Maybe for example in your use case, how you're using Zapier. Maybe we can add that integration into the platform. Some of the things that maybe we're looking to add soonish is the ability to hook it up to slack, hook it up to forms, because we noticed a lot of people, they'll have a lot of capture lead capture forms, and then what they do is they'll send that data to otto grid to enrich maybe all the details and then send them on slack, a notification if the users qualify.

41:46
sully
Just an example, but I'd love to learn more about how your automation would work and we can directly add it to the platform, whether that's a zapier or make integration.

42:00
Ben Tossell
So we have people asking, obviously they want to get access.

42:09
sully
You said email you, or we could set something up. Maybe I can throw my email in. But anyone on this call, if they want access, just. I'm more than happy to give you guys access to autogrid.

42:23
Ben Tossell
Awesome. That's an exclusive right there. And then I think Gurpreet asked a question that I think I answered. He basically asked if you can upload transcripts and it generates a table and it'll pull out like people books or products mentioned in it. I think the example you gave with the call summaries is that similar. Yes.

42:44
sully
So here you can actually see features requested, summary use case problems faced. Yeah. So you can do that as actually, again, one of the nice parts is like this is sort of like those workflows where like I, I grab the name of the person from the call, grab their LinkedIn profile, and then I'm like, grab me their job title and like what they do based on their LinkedIn profile. So you're sort of creating like these, like get their workflows right if you kind of think about it, but you're able to like do it directly in a table and you can see the data like as if it's in a spreadsheet. Yes.

43:20
Ben Tossell
And then we've had a feature request for chrome extensions. You can get behind payrolls. And. Yeah, people have been asking about the web search agent, like, what is the best web search agent out there? What do you guys use to do your web search? Was it?

43:37
sully
Yeah, great question. So we spent a lot of time and effort the last year building our own web searching agents. We're going to release a benchmark soon, but basically our new one that's not live is the most accurate web searching agent that's available. We compared it and ran against perplexity. We ran it against cohere and chat GPT. We basically built our own in house web scraping agent, our own ways of making sure that the data that you source is accurate. All of it is built in house by us. We also do provide more data. So I didn't show you guys, but you can add more data sources if you're like, I need a list of emails or I need to know specific information.

44:21
sully
Like some people have Zoom info, some people have clear bit all these other data providers where they're like, I'm already paying for it. I'd love to just be able to pull it in. We do let you do that. Our first one is Apollo, but in terms of just web searching agent, we have our own one that runs within our own scrapers and our own whatever agent module you want to refresh to it as. But yeah, for those of you who want to please shoot me an email because I'm going to forget if you guys have it in the chat, just send me an email saying, hey, this is the email. One other thing, for those of you who would like access, if you could go in and create an account, because to give you access, I need you to create an account.

45:00
sully
But other than that, just shoot me an email, I'll get you guys all access. So I posted it here, I'll post it again. Oh, I guess it's a direct message here. Actually I keep to everyone in this meeting. Sorry. So yeah, it's cognosis here. There you guys go. I sent it directly to red and I was responding to a single person. There is my email in the chat. Shoot me an email if you guys would love access. Excited. We're still in like a beta, so I would say, you know, if something goes wrong, please let me know. We'll be launching soon. We'll be launching very soon. But there's just some issues with like scaling, trying to have like you know, hundreds of thousands of agents. It was a little challenging to figure out, so, but we'll be launching the product soon.

45:43
sully
You guys will have access if you don't want to join the waitlist.

45:45
Ben Tossell
So awesome. Well, yeah, we really appreciate your time and thanks for doing those demos. I think a lot of people are really interested in autogrid, especially by how many people are pinging fraxis. So yeah, have fun dealing with those emails. But yeah, I mean this has been great and yeah, really appreciate your time and we'll have the recording up on the site on your event page so people can watch it afterwards and yeah, hopefully we can have you on again when it's all launched and you've done even more agents and weird wonderful things there. That's been great. So yeah, thanks for everyone for joining.

46:25
sully
Sorry, I didn't mean to cut you off there.

46:27
Ben Tossell
Yeah, yeah, if. Yeah, if you're happy with sharing them with. We're happy to share them around. I know they just sort of examples.

46:35
sully
How useful it would be but here I'll just save it and then copy anyone link.

46:44
Ben Tossell
Great, yeah we'll put those on the event page as well so yeah people can get the access to them. Awesome. Well yeah, I appreciate your time and yeah, cheers everyone and we'll thank you around next week for two more expert workshops. So yeah, more coming. Cheers man, cheers.

47:05
sully
The recording.

‍