Automating work and streamlining productivity with Gumloop

Max Brodeur-Urbas
Yeah, I'm Max. I built Gumloop with my co founder, Rahul. We definitely should be a bigger team, and hopefully when you see the product, it'll seem like it was built by a bigger team. But we've just been in the process of hiring so far. But we started gum loop about a year and four months ago now. And it all started by me building the first UI for auto GPG, which was like the fastest growing open source agent framework. And what I found when I joined the auto GP community on discord was that half the people in there wanted to use agents. They wanted AI to do work for them, but they didn't know how because auto GPT was an open source project. You had to clone it, set it up locally. So I just built a front end for it.

Max Brodeur-Urbas
I learned how to use react and pieced that together. I called it agent hub at the time. And when someone would have a technical question about auto DBT, I would send them a link to Agent Hub. And that's how we got our first users. And what we learned by talking to those users was that 90% of what they wanted to solve could have been solved with like a python script, a couple API calls and some GPT. Requests, but they were just hoping that the agent could handle the work for them. Agents were disappointing them pretty much every time they tried to use them. So we would get the complaints people thinking we built auto GPT when we didn't. So we came up with a more rigid step by step framework.

Max Brodeur-Urbas
And that just naturally evolved into what it is today, which is we get compared a lot to Zapier, but I think it's a lot more complex and hopefully can automate much more complex work than Zapier, but we integrate a lot more AI natively, have a lot of the same, and ideally you can fully automate really intricate workflows for your business on gum loop.

Ben Tossell
Awesome. Yeah, I think I've heard the Zapier for AI workflow. I mean, I saw a lot of those kind of companies pegging themselves like that. But yeah, I think having played around with the product, there's so much more you can do. And like, all the AI native stuff does really make a difference when it's built as an AI native platform. Right. So, yeah. Excited for you to show us some of those things.

Max Brodeur-Urbas
Yeah, I didn't know what Zapier was when I started all of this, which I'm proud of. It was not trying to be this, but I guess it just ended up being the. I think it's the shortest path to extracting actual value out of LLMs when you can constrain them and have them perform specific tasks. And one counterintuitive thing we do on Gumloop is we encourage people to use as little AI as possible because that drives the cost down, increases reliability, and you can normally get 90% of the work done with just data manipulation, calling APIs, writing and pulling data from sources.

Ben Tossell
Yeah, that's interesting. Yeah, I think because people, potentially, I'm thinking same people, I'm thinking of me when I first started building stuff. And it would have been around. Zapier is looking for the AI to do the work of the data extraction and all that kind of stuff. But actually it might not have needed to be LLM doing that work, especially when they're not so deterministic with what their outputs are. And it changes all the time. So it's like figuring those things out, which I think. Yeah, really interesting to see that. So I'll stop interrupting and let you go for it.

Max Brodeur-Urbas
Cool. Okay.

Max Brodeur-Urbas
This is. You can see my screen.

Max Brodeur-Urbas
Yeah.

Max Brodeur-Urbas
Okay. This is what gum loop looks like today. This is my hub. We used to be called Agent Hub, like I mentioned, but everyone thought I was saying asian hub when I would speak too quickly. So we rebranded, we also got the.com for much cheaper. But yeah, this is the hub where you see all of these would be the equivalent of your zaps. I'll go ahead and make a new flow just to show you how you go about building something. And then I have ten use cases I could run through. I'll probably stop after three, most likely. So you start off with this big empty canvas, and the idea is you use these nodes, we call them, which are Lego blocks, and you connect them together on the screen. So let's say I wanted to read a file from Google Drive.

Max Brodeur-Urbas
I would go into our data loaders and grab a Google Drive file reader. And then I wanted to use Claude 3.5 sonnet to extract the byline or the first name on this report card. This is just a contrived example, but I would go and pick Claude 3.5 sonnet, use this data extraction node, and specify the schema like first name, and I would explain what that is down here. And then you can contextually build this data extraction node. Our AI nodes are kind of our bread and butter. They let you leverage LLMs without having to worry about what's going on under the hood, how you get reliable sort of output. So if I wanted to do like grade, I would now have these two values to use later in my workflow, first name and grade.

Max Brodeur-Urbas
But you can summarize data, categorize it with natural language, you just define the categories. You can score your data, you can analyze PDF's with vision models, analyze images, you can fill PDF's. This is a feature that is insanely intricate, but we don't really advertise. So yeah, these AI nodes let you leverage LLMs, and then how you connect data into your flows is where the real value is when you pair AI with your own data. So reading from your Google Docs, scraping the web, your gmail, outlook, Reddit, YouTube Typeform notion, we continuously build these integrations so that more people can tackle the use cases. And all of these nodes are driven by consumer requests. So when a customer has a request for a node we don't support, we like to say we have same day shipping, so we'll get the node out same day.

Max Brodeur-Urbas
But that's not totally true anymore, just because we have to put too many requests, but we try to stick to it. So I will show you a flow that's actually useful. This one generates content for a YouTuber who wanted to convert his YouTube catalog of tech reviews to SEO optimized blog posts. In this flow, I'll just run it to show you what the output is, and then I'll explain how it works. It takes in a link to a YouTube video as input. So the start of this flow is just an input node where he puts in a link to his YouTube video, he presses run, and then these two nodes, which are subflows, I'll explain those in a second, get executed, and the blog is live within seconds here.

Max Brodeur-Urbas
And it's formatted, the tone is corrected, it has all the observations you mentioned in the tech review, plus the video embedded at the bottom of the screen, and it was published directly on our CMS, which is Ghost. But this whole workflow is just a series of strung together AI requests that are prompted to convert the YouTube transcript to a thoughtful blog post. So what I just did here, this node is actually an entire separate flow that's a function on Gumloop. It's how you compartmentalize functionality. You keep it really modular. It lets you test individual actions, and then when you piece them together in the larger flow, you get much more intricate automations. But this generate blog post subflow is an entire flow in itself. This is where all the magic is happening in terms of going from YouTube video to thoughtful blog post content.

Max Brodeur-Urbas
It looks a little daunting, but if I zoom in, it's as if you prompted GPT ten times in a row and we're copy pasting prompts between your chat windows to get to your desired output. The input here is the YouTube video. Whatever we pass in dynamically, we then get the YouTube transcript with this node that gets the transcript. And if you were to try to generate a blog post by just asking GPT to go from transcript to blog post, it'll start hallucinating, it'll underperform, it won't listen to your instructions or any of the nuanced things you asked it to do, like the tone or I the verbiage you wanted it to use.

Max Brodeur-Urbas
But if you split this up into three different prompts, you can let it focus on the introduction, or the observations in the middle, or the conclusion at the end, and then piecing those together makes up creates a much better blog post. So here I ask it to generate an introduction. I say speak in the first person past tense, and all this GPT 400 request is doing is the introduction. Same sort of thing for observations. I want it to tell me everything that the youtuber observed about this tech review. Like the pros, the cons, the in the middles, and then a conclusion as well. I format all of that back together and then I run it through a tone checking step.

Max Brodeur-Urbas
So one problem we noticed when were generating the blog post originally was that it would throw in words that didn't match the youtuber's tone. So I append an example snippet of how this youtuber speaks and I ask it to only make small changes, see if you can improve the tone. And then here I format an HTML so it takes the blog post and just reformats it. And then the output is a really high quality blog post with these headers specified. That's why everything is bolded and indented properly, because the last AI step formatted it in HTML. And the loop part of Gumloop is where the real magic is. That's where you automate things at scale. So just like this node can be, this node is an entire flow. This whole flow can be used as a node as well.

Max Brodeur-Urbas
So you can infinitely build upon your subflows. And what I would go to is this subflow over here called publish at scale. This entire published blog flow is being ran on all the rows in a Google sheet. So if you have 1000 YouTube videos in a Google sheet, you can loop this entire published blog flow a thousand times. Just like if you had a function that was good to write one blog post, you can just put it in the for loop and run it at massive scale. And gumloop handles all the infrastructure of how you actually scale your operations. We have parallel step execution, for example. So this will actually do 15 concurrently at a time if you're on the pro plan. And yeah, it can run for 1010s of hours.

Ben Tossell
Can I just ask a few questions that are coming up in the chat? So with that example I just got. My question first is with the Google sheets, can it also, does it also get triggered every time there's a new link put in the sheet? That's how a zap would be a trigger. When the new row is added to the sheet, does it work that way as well?

Max Brodeur-Urbas
We have a group of triggers. We have things for Gmail and slack and calendly and airtable. The airtable one works exactly as you described. For airtable. We haven't added the Google sheets one, but it's something that we could add within a week most likely. Cool.

Ben Tossell
Awesome. So then a couple of questions from the group. Can you have a human in the loop in the workflow and edit the blog post without interrupting the flow before publishing?

Max Brodeur-Urbas
Not directly on gum loop, you can have build your own human in the loop experience. Dump it into a Google Doc instead. If I wanted to not publish it directly, but dump it into Google Doc for review, I would just swap out this node and then review it and then have a separate subflow that goes from a Google Doc link to my post. So you engineer your own human in the loop experience?

Ben Tossell
Oh yeah, I see. Yeah. So you save it as Google Doc, and then you could say, when this gets added to a folder in Google Drive, then.

Max Brodeur-Urbas
Exactly.

Ben Tossell
Awesome. And then with that example here, this one. How much does it cost for that workflow? How much credit does this work spend to process a generic radio?

Max Brodeur-Urbas
Yeah, good question. So you kind of build your own cost. It varies quite a bit, depending on how you build your workflow. So running this flow itself costs one credit. The basic plan comes with 30,000 credits, but there's an AI cost associated. So this flow itself costs one. But if we go into this blog post generator, each one of these queries costs. I mean, the credit cost is associated with the model that you're using. So if you're using a state of the art model like GPT 4.0 or cloud 3.5 sonnet, it would be 20 credits per AI request. And if you were to downgrade it to GPT 400 mini, it would be two credits. If you were to provide your own API key, all of these would be one credit. So you wipe away the cost.

Max Brodeur-Urbas
If you want to just throw in your own API key.

Ben Tossell
Okay, yeah, that's good. Yeah, awesome.

Max Brodeur-Urbas
Cool.

Ben Tossell
Yeah, we can move on.

Max Brodeur-Urbas
So this flow, I think, is cool to show off because this is how we generated most of our revenue during Y combinator. In the recent batch, our problem was that people wanted to use gum loop they were signing up, and we made it very easy to sign up, but weren't really able to give them that aha. Moment of knowing how awesome the platform was. So I'd want to get on a demo call with them and show them. And this flow is what set up all of my demo calls. So the input to this flow is just someone's email. That's all that we had when they signed up. And what I would do manually before I made this flow was I'd go research everything I could about them and figure out if they were a qualified lead in terms of like, a potential enterprise customer.

Max Brodeur-Urbas
But what we do now is we research their company with Gumloop. And I don't sit here copy pasting emails every day when people sign up, I trigger this via a webhook, which is a slightly more technical feature. But basically the second that they press create an account, this flow gets triggered like an API, and it runs in the background doing the research. So their email dynamically gets passed into this flow as input. Then here is where the real magic is happening in terms of the research. If I go to view this subflow, it is grabbing their email and getting the domain out of it. If it's Tim at HubSpot, it's slicing the Tim at part out and just getting the domain, which is what these two nodes are doing. Then it does a few things with that domain.

Max Brodeur-Urbas
One is it scrapes their website and then summarizes what they do with cloud three haiku. But you can swap any model in. You can also modify the prompt template under the hood, which is a more technical feature. But if you want more control over how things are being summarized, you can do that. I extract the company name. This was a modification I made down the road, because if you try to get a company name with a data enricher like Apollo, you might get Gumloop YCW 24, because that's what we have on LinkedIn. But that's not how you would refer to our company.

Max Brodeur-Urbas
So I thought it would better for AI to just look at the website, tell me how they refer to themselves, and then I do use those services like Apollo, Proxy, Curl, similar web to enrich this information, I get their industry, annual revenue country. I could be getting their monthly web traffic, job postings, total funding, LinkedIn. But all I really cared about was the revenue as a proxy for how big of a customer they might be. Then I format all of that information in this notification, and then that notification gets sent to research results as an output. And you'll notice in the parent flow, research results goes directly to our slack channel.

Max Brodeur-Urbas
So we have an enterprise signups channel, and basically 10 seconds after they press create an account, we get a full research report on how big their company is, how much revenue they're generating, where they're based, a summary of their website, everything we really need to know. And that's how I know to reach out, probably later in the day. But the reach out is also automated. So I have this subflow that's a really rigid GPT 4.0 prompt, and I'm giving it information about the company saying personalize it, and say I'm the founder. Here's my meeting link that gets dumped into my Gmail as a draft, so you could send it off right away. But it's better to not send a personalized email 10 seconds after they sign up.

Max Brodeur-Urbas
That was how every night I would send off maybe 20 of the most exciting leads, hopefully get on the call with ten and convert a handful and generate revenue.

Ben Tossell
Thats awesome with something like this. Kevin has asked about how easy are these sets up because obviously looking at this one you think ok, theres four or five nodes there, thats great, I could get that built and set up. But if you are new to something like this, develop those sub flows and then there's obviously plenty of steps within that. And then even I think in the user email company research subflow there was the one bit where you said this then connects to the research results. And to me that wasn't, it wasn't like an obvious oh this is the thing that then will pull out the available connector on this one.

Ben Tossell
So how do you think about all those different node types and like understanding it so that a newbie or a really non technical person can come in and say I can build like someone seeing this flow feels like I should definitely have this flow and it could be tweaked to whatever example, but it feels like an obviously oh you should definitely have something like this. But how easy is it to build?

Max Brodeur-Urbas
Yeah, the learning curve is definitely steep and that's like our biggest problem at the moment. That's why I had to get on demo calls with all these customers and build this flow. Once someone invests like 30 to 40 minutes tinkering around, watching the tutorials that we have on the hub, which I'll show you in a second, and starting with templates which we have pre built, then it's much easier. But yeah, there are some UX details that are not super intuitive like the way you mentioned like research results is coming out here. You don't know how you defined it. You have to have an output node. So you use input and output nodes to define inputs and outputs for your subflows. That's not totally obvious to someone who hasn't seen these before.

Max Brodeur-Urbas
So I think one of the unlocks will be when we use AI to help you build. And you don't actually have to discover the nodes yourself. We're working on that. But one thing we know is that will be the bottleneck of the quality of the product. No one will ever build manually again if we release an AI builder. So someone might not think gum loop is powerful because the AI builder wasn't good enough. So we need to make sure it's really accurate and powerful.

Ben Tossell
Yeah, it makes sense. Yeah. Something I saw yesterday on the product I was testing that is when there was an AI copilot assistant, when there was an issue, I was sort of saying oh, I'm trying to do this, or this doesn't work. It actually screenshotted the screen, like actually did a screenshot of. So you can see the whole thing and be like ok, I can see that this person is trying to set up a cold outreach or like a research agent and they're missing this output step or something like that.

Max Brodeur-Urbas
That's cool.

Ben Tossell
That's kind of thing that you'd need without pinging you or your co founder to be like hey, I can't figure this out.

Max Brodeur-Urbas
Yeah, we built this thing called Gummy here, which is a support bot that knows all the nodes and can help you build a. This thing will hallucinate like crazy. We shipped it in like 12 hours. So it's not the most thoroughly built system, but it does help. But yeah, quality is not 100%. I can walk through another use case that I think is really awesome. This one, like the SEO automation space, is something I'm pretty new to, but as I've been diving into it, I found some really useful automations. So this one takes a grouping of related blog posts. It then tries to come up with, it tries to find all of the keywords that whatever company wrote this blog post is missing or keywords that they should be hitting in the related area that they've already been writing about.

Max Brodeur-Urbas
So what it does is it takes all the blog posts you specify, it runs this upload that's really simple. It just goes blog post by blog post, scraping and summarizing each one to determine what they were talking about. Then with all of those summaries, I ask AI to come up with 60 ideas for keyword high intent search phrases that someone might be looking for in this sphere of topics. And then it runs each of those keywords through Semrush to get the search volume and keyword difficulty. Now it's armed with all of the data to know what keyword's valuable, what are the low hanging fruit? And then I feed all 60 of those reports of keyword difficulty and volume into a final AI step where I ask it to tell me which ones I should focus on. What are the low hanging fruit?

Max Brodeur-Urbas
So if you run this, it writes a Google Doc that's formatted with HTML and it should be a report specifying the top keywords. So you can modify this to go post by post. You can modify it to actually go through your entire website if you wanted, and read every single blog post and do like thousands of keyword searches and write a 20 page report. But this is what I put together in like 15 minutes. Oh, this is not that good. The last one that I did was so you can debug all of these by looking at the inputs and outputs and tweak the prompts. Let me tweak this. So be thorough in your reporting and explain why keywords are valuable. We can resume this step from here.

Max Brodeur-Urbas
So if you ever want to rerun a flow from the last step, you can debug it and resume it from that last step. And then that lets you quickly iterate on your prompts, tweak the outputs. You can also like iterate on the model that you've picked by just swapping it out and rerunning it from that exact step. So it's taking longer because it's writing more output in terms of the analysis. Okay, something has gone seriously wrong. I think it didn't find any valuable keywords.

Ben Tossell
The curse of the live demo again. I mean, it's quite interesting to see you run them and something not working and then looking at ways to debug. So I think there's a few challenges with something that is this powerful, which is you've got to know the right nodes and how to connect them. And then, like you just said, you just have to iterate on your prompts. That's another challenge. And then there's, sometimes things don't work either. So, like, how do you actually go around and fix something that feels like it should be working?

Max Brodeur-Urbas
Yeah, it's interesting. It got like 50 of the values. We can see all of the search volumes and the difficulties for all of these keywords that recommended like federal income tax calculator. That one has 22,000 monthly visits, but a keyword difficulty of 100. And it seems like the data being passed into the AI model is great, but what's going on here is just not what you would expect. So I swapped out the model, but yeah, it should be rigidly defined enough where you can investigate each step and not ever wonder what's going wrong. The fact that I'm wondering what's going on here is ironic, but yeah, I'm just gonna wait. I'm just so curious why this is not working. So I need to watch. Yes, definitely watch it.

Ben Tossell
I think we're all wanting to see that too. Dave's asked, can you use keywords everywhere instead of Semrush for keyword research? Can you do that with different things or does it have to be a specific note for that.

Max Brodeur-Urbas
Yeah, so the nodes just make it a lot easier to do this sort of thing. Like it integrates directly. This is broken. Okay, this is just terrible. It wrote, okay, I'll stop here. You can call any API directly. So if you wanted to push the boundaries of like we don't support a node, but you wanted to call the raw API with your auth key, you can pull in that data directly. You can also execute arbitrary python code, which is another more advanced feature, or JavaScript code if you wanted to write your own functions. And a feature that we're coming up with in the next couple of weeks is the ability to ship your own custom node.

Max Brodeur-Urbas
So you can define the inputs, define the outputs, and then basically create that reusable block in a node builder and you either ship it for your, use it yourself or you build it for your team. But yeah, so in short, we could add an integration for that service that you're talking about, or you could just call the data directly in the meantime.

Ben Tossell
Yeah, this node and then yeah, around that Semrush step was the error shield thing. What was that? John's asking what that is.

Max Brodeur-Urbas
That's a great question. Okay, this is what the doc should look like. This is what, right before the demo, what I ran before, I tweaked it to use subflows, but it just sorts it by most valuable based on keyword difficulty and volume. An error shield is a yemenite, definitely a more advanced feature, but it is a way to avoid your flow from crashing if something were to go wrong. So it's basically a try catch if you're familiar with that concept from programming. Oftentimes for web scraping, for example, when we're doing that research flow that finds information about someone's company and scrapes it. Some people's emails don't have a valid website associated, they just have the domain, they use the email, but there's a 404 error. So the error shield lets you skip that scraping step and just proceed gracefully.

Max Brodeur-Urbas
Got it.

Max Brodeur-Urbas
This is the series of tutorials I was talking about on the hub. They're just loom videos that you can walk through and it should be confetti when you complete it. But yeah, there's also a link to our YouTube channel and Discord community. There's about 1100 people in the discord that are asking questions, getting free support. So if you ever want to tinker around with the plan, with the product, you can do that on the free plan without putting in a credit card and get free support in the discord.

Ben Tossell
Awesome. We've done a couple of tutorials, I think, and have more coming out as well, so hopefully we can make it simple to follow some of these as well.

Max Brodeur-Urbas
Cool.

Max Brodeur-Urbas
This one, I think is really cool as well. It does a lot of LinkedIn scraping and email processing, and uses triggers. It was originally built for an investor who wanted to process incoming pitch decks. They had analyst that was just receiving pitch decks and they had to read them, like open them whenever someone forwarded it to the analyst inbox, extract any information, and then enrich their database with this new pitch deck that they've received. But what this flow does is it starts with a Gmail reader that reads the attachments and the email bodies from an incoming email, and this whole flow is triggered with a Gmail trigger. So what they did is they created a new inbox called pitcheshervcname.com dot.

Max Brodeur-Urbas
Whenever a partner or analyst received a pitch deck, they would just forward it to pitches and it would trigger this flow immediately when they received this email. What it does is it extracts, it takes the PDF attachment, which is the pitch deck, and it uses GPT four O vision to go page by page and extract all the content. We do have a normal PDF reader that just uses basic PDF libraries, but with content like this, a lot of it is very visual, wouldn't get picked up by a normal PDF reader. So vision is good for resolving those graphs or charts embedded in the PDF's. I then format all of the pages from the PDF, whether it's five or ten pages, into one big chunk of text. And I have three AI steps run on it. One extracts the key metrics in point form.

Max Brodeur-Urbas
I encode a little bit of nuance here, where I say, if you see massive numbers, sometimes a slide will say 15 billion with no context. And you can tell the model that's definitely a projection. It's not a real metric that they have. Unless maybe you're like a Series F investor, then that would be relevant. But this was a seed stage investor. So same thing here. I say extract the company name and revenue, but if you don't see a revenue explicitly stated, you can just say pre revenue. And then here's the interesting part, I find it gets a list of founder names from the pitch deck, and then for each founder name, it runs it through a subflow with an error shield, so that if this were to go wrong, we can just skip over it.

Max Brodeur-Urbas
It finds the founder on LinkedIn, scrapes their work history, and summarizes basically everything they've done in the last, I mean, the recent ten years of work. And it writes all of that information to a notion database and then sends them a link in their new pitches channel that says, here's a link to the new notion entry. So the entire flow is just forwarding an email and then you receive a slack ping like 30 seconds later. With the nuanced research and extraction from that pitch deck, I can show you the LinkedIn subflow, because it's not using LinkedIn search directly, it's just using a cheeky workaround, which is something you could build into your own agents, if you guys are building agents. But it's utilizing Google search in a particular way to get LinkedIn profiles accurately.

Max Brodeur-Urbas
So all that is doing is it takes the founder name and the company name and then it formats it in this specific way. So it would be founder name, company name. And then I specify site, LinkedIn.com, in, which means that Google will only return results that follow that format that are LinkedIn.com in which is a profile. I then use our web search node, which is like, I mean our Google search node. Essentially we do have advanced options that you can do, like Google Hotels, events, news, if that's what you wanted to scrape with the SErP API. And then I grabbed the first profile, which should be their profile if Google has done its job correctly and if they're like a non stealth founder. Basically I scrape their website and I extract a list of data.

Max Brodeur-Urbas
So I'm defining a schema here for, I want job title, company name and duration for every company that they've worked at on their profile. And then all of that is dumped into this large chunk of text where I ask AI to just give me a two sentence summary about their background. And then I format the link at the end and I dump that into work history. And that's what gets dumped into founder backgrounds in the notion.

Ben Tossell
Awesome. Do you have any examples of what the end looks like for this?

Max Brodeur-Urbas
Yeah, I, I can run it live based on the last one. You would think this wouldn't work, but. Okay, I'm going to drag this here. So this was my test folder. It was the intercom pitch deck. Kind of hard to find startups, actual pitch decks from their early days. But yeah, this is one I found. And then I have this links into pitches. So I press run it, will read the email, it has to convert the PDF to an image and goes page by page. But yeah, now it's extracting name and revenue of the company. They're pre revenue at the time and then it's searching all three founders. On LinkedIn, it's doing all three at the same time because we have stephen parallel execution. So three founder names got fed in here and it's doing all three at the exact same time.

Ben Tossell
Do you have to do anything to specify by doing that step function thing? Shed.

Max Brodeur-Urbas
No. On the free plan you get two at the same time. On the starter plan you get six, I believe, at the same time. And then on the pro plan you get 15. Some of our enterprise users have up to 30 concurrent steps being executed just because they're doing tens of thousands steps a day. So the notion DB link is here. So Intercom, I don't know how to pronounce his first name, but McCabe, based in SF, his entire work history is literally just intercom, I believe. Yeah, basically. So not much to summarize there, but.

Max Brodeur-Urbas
This.

Max Brodeur-Urbas
Other co founder is here and it just found them by name by like Googling as you would. And to use that kind of cheeky little Google little trick. And then the projected metrics. Yeah, they had basically no traction at the time. If you read through their pitch deck, it was just an idea and a couple of really qualified founders who wanted to build. Yeah, we also got a slack message summarizing the pitch deck in our channel.

Max Brodeur-Urbas
Nice.

Max Brodeur-Urbas
Well, I can do one more use case that people have been liking. This one analyzes my calendar every day. So it researches all the people I'm going to meet in a given day and then sends me a text message with a summary of those people I will encounter. And then an email which is a more like verbose summary of everyone's traction. Basically doing that sort of research flow that I was doing on a customer by customer basis. But every morning at 08:00 a.m. And then dumping that research to me. So it reads from my Google calendar, I have the relative dates, so like today and then 24 hours from now. And then if I were to run this 08:00 a.m. It would do the next 24 hours of meetings. It gets the attendee emails and runs a subflow which does research.

Max Brodeur-Urbas
And then it has three separate AI steps. One encodes all of the nuance saying like I'm Max from Gumloop. If you see gumloop meetings, assume those are internal meetings. Don't put too much thought into those. And I want you to write a concise report about who I'm meeting today, but focus on the people with the most web traffic, the biggest company size, highest revenue. And then I have two AI steps, one that prepares it for the text message. So going from that large report to the TLDR and then another one that goes from that large report to a table format so that in my inbox I get like a nicely formatted table that separates morning and afternoon and all the metrics about the people I'm going to meet.

Max Brodeur-Urbas
And all you would have to do is like enter your email and your phone number and choose your calendar you want it to read. And this would be scheduled with a time based trigger to run like every morning at 08:00 a.m. That's cool.

Ben Tossell
So how many of these flows do you have running for doing actual jobs to do work? So whether it's the enterprise flow daily calendar one, how many do you actually have running?

Max Brodeur-Urbas
Internally at Gumloop, we have probably six to seven going every day. One at 04:00 p.m. Calls our internal API, getting all the metrics from the day and telling us who the most frequent free users are, who the most active power users are, and sends us an HTML formatted report in our inbox to myself, my co founder, and our intern. So we all have an idea of today's traction and then the research flows. Every time someone subscribes on stripe, we get a notification that finds their LinkedIn and tells us their position at their current company that in slack. So half of my pings normally throughout a given day are AI generated research about traction or customers.

Ben Tossell
And then have you thought about building flows to analyze the pings? Because I assume the more automations you end up building to take away a load of the other work also gives you a lot of these notifications and then like how to manage those, I guess like is there a way to do that with this?

Max Brodeur-Urbas
Yeah, I think there's a way to maybe reduce them into more important notifications. So for this enterprise flow, we used to be trying to find every user that signed up, but now we just do people who have non Gmail outlook accounts. So like company emails, that gives us a lot more information. And then the next step after that would be to filter out any company that doesn't have like over a million ARR or something if we're trying to like really bucket people. And all we have to do in that case is add an if else node in our research flow and then you can basically split the paths of there we go. Okay, you can add conditions in that research flow.

Max Brodeur-Urbas
I could have a condition that says, if this is not above x value, I want you to pursue this failure flow if it's above this value, then notify me. So maybe I could go from 200 notifications a day to 50 really high quality ones.

Ben Tossell
Yeah, it makes sense. Also, man, we can open up the floor to questions if people have them and they want to raise their hands using Zoom. Obviously Jeremy asked a question. I don't know if he is willing to come on live to ask it. Just I want to make sure that the question gets asked correctly.

Jeremy M Phillips (hehim)
Yeah, a lot of my common automation needs in my company are for interacting with web applications that don't provide an API. And it's not just scraping data from web applications, it is going into web applications and submitting forms. Often it's moving data from one web application that doesn't have an API to another web application that doesn't have an API. And I'm just curious if Gumloop supports those use cases or if you have future plans to do so.

Max Brodeur-Urbas
Yes, that's a common request. Our scraper works wonderfully for getting content that's publicly available, but if you need to log in then that throws a wrench in the system. But one thing we're working on now is a gum loop chrome extension which will let you trigger any flow and capture the content of your page from the context in which you're seeing it. So like behind the login or behind the paywall if you wanted to do that manually. And it will also let you record actions on your screen as replayable steps within your flow. So if you wanted to click type in username, click type in password and press enter, you wouldn't have to do any coding or scripting to do that. You would just record yourself doing it and then that is a replayable action within your flows.

Max Brodeur-Urbas
Awesome.

Jeremy M Phillips (hehim)
Thank you so much.

Max Brodeur-Urbas
Cool.

Ben Tossell
So that's more like the RPA kind of scenario with recording a screen and then mapping that to a flow.

Max Brodeur-Urbas
Exactly. Yeah, I think it would unlock a few thousand more use cases in terms of the data that can be pulled in. But yeah, I think the ETA on that is like two weeks. We are still two people and doing a lot of interviews and stuff to hire, so we have to find time to code. But yeah, that should be out in a couple weeks.

Max Brodeur-Urbas
Awesome.

Ben Tossell
Yeah. Anyone else with questions please raise your hand. I'll go through some in the chat. So may Rose has asked, is it possible to populate a canva template with some variables that are text? That is a sheet plus canva equals output.

Max Brodeur-Urbas
So you want to have a flow that has variables you can pass in dynamically.

Ben Tossell
I think so I think it would be like if you had a blog social share image template on canva and then you wanted to change the. So for this event we had the name populated as a variable and the image populated as a variable. And that all comes from a Google sheet is what I'm assuming. So Mary Rose, let me know if that's wrong, but yeah, totally.

Max Brodeur-Urbas
The input nodes let you define dynamic variables. So you could specify the candle, link, the title and the image separately as variables and either pass those in as part of the flow. There's also another thing I didn't really show you guys yet, which is called our user view. Let me go to this other flow. It lets you deploy these flows as basically apps. You could pass in these variables dynamically. So this flow generates a cover letter. One dynamic variable is a link to the job posting. Another one is my resume as a file. And we can deploy this flow as what we call a user view. It's going to be rebranded to gum loop forms, but this is a way to view your flow in a more interpretable way and see the variables I guess you wanted to pass in.

Max Brodeur-Urbas
It could just be the three variables you care about and the file or the image you want to. What is the header?

Ben Tossell
Awesome. Yeah, I was going to ask if there's going to be some sort of front facing UI that you could have for that instead of, I guess, going into those flows. It looks like it's this.

Max Brodeur-Urbas
Yes, I built this six months ago and never advertised it. A lot of our power users embed this in their websites and use it all the time, but I don't do a good job of showing people it exists.

Ben Tossell
Awesome.

Max Brodeur-Urbas
Yeah.

Ben Tossell
Any other questions? Pop them in the chat or raise a hand and we'll get you on to ask them. Jeff asked earlier maybe this is similar to Jeremy's question as well, which is, are there certain websites that are impossible or difficult to scrape using gumloop? Obviously the logging in version is very difficult, but yeah. Are there any others that generally are hard to do?

Max Brodeur-Urbas
LinkedIn is pretty notorious for being anti scrape. We make sure not to ever like log into a LinkedIn profile or try to get around more than just what's publicly available. But there are a lot of websites like that where there their content is the product. So like New York Times or something like that is harder to scrape than just like a normal like job board in like Canada. You know, like they have different levels of care in terms of how they protect their information.

Ben Tossell
Yeah, makes sense. Cam has a question if you want to jump on and ask Max, this is amazing.

Max Brodeur-Urbas
Thanks for sharing this with us.

Jeremy M Phillips (hehim)
I'm curious, have you found any patterns?

Max Brodeur-Urbas
Patterns and how people figure out what to build with gumloop? It seems like such a powerful tool that could be a little bit overwhelming sometimes. How do people find what to build and prioritize that? I'm curious if you've seen anything. Yeah, I've definitely noticed that there's a failure of imagination that happens even with myself when I just go to gum loop and decide I'm going to build something. It's much easier to start with a concrete problem that you have in your business and then go work backwards and automate that with Gumloop. Then from there people get really familiar because they tackle their first use case, something that's a burning need for their business, and then they feel comfortable exploring tinkering with other use cases. And it spreads laterally like that. We do have a templates directory as well.

Max Brodeur-Urbas
I spent a lot of time making these. It's a dynamic directory of different use cases that you can tackle, and they're all just pre built like this one is automated invoice, classification and extraction. I try to convert everything interesting that I built into a template, and I actually do a pretty poor job of it because half of the things you saw here aren't templates yet, but they should be in the near future.

Max Brodeur-Urbas
Thanks.

Ben Tossell
There should be some sort of flow that helps generate these templates, surely?

Max Brodeur-Urbas
Yes, there should be. We have a flow that writes all of our documentation. Actually, all of these pages are AI generated with a gum loop flow, but we should apply the same logic to the templates for sure.

Ben Tossell
Yeah. Awesome. John, under Jeff, you have your hand raised. Do you want to ask a question?

Jeff Kaminsky
Yeah, I'm curious first, since you just mentioned that one, did you open source that template for creating documentation? And then two is, I've been playing a lot with Lindy and they're coming up with a workshop soon. I'm curious how you position yourself relative to them just so that I can have an understanding of which service I might use.

Max Brodeur-Urbas
Yeah, good question. I've actually never used Lindy. I've never used most competitors. That's kind of something that we try intentionally to do, because we find if we use something in the space that's tackling a similar problem, we'll get stuck in the paradigms that they decided on and think those are right without, like, coming up with our own ideas based on just like, first principles. Completely fair. So I've actually no idea what Lyndy does, to be honest. I've heard of them and I'm sure their product is awesome, but we're just focused on building what our current users are asking for and showing other people how cool the product is.

Jeff Kaminsky
And open sourcing the template for creating documentation. Is that one of the templates?

Max Brodeur-Urbas
It's not one of the templates because it's committing directly to our GitHub repository. Interesting. Yeah, but it's really just like we generate the docs in Markdown, so it's just reading a Python file, which is like the schema defined for an integration, and then I have a lot of prompts going through it, trying to extract the nuance of how you would use this scenarios for which it would be useful, and then it converts that to markdown and publishes it to GitHub.

Jeff Kaminsky
It sounds amazing. It also sounds expensive. How much is it to run that process?

Max Brodeur-Urbas
It's a good question. Our bill, it's the drop in the bucket in terms of, for us running that workflow compared to just like all of gumloops AI costs. So I haven't actually measured that exact one, but it's probably around like 100 credits because it's five GPT 4.0 calls. Maybe one of them is a GPT 4.0 mini call to format it to markdown because it's really easy. But a pattern I notice is that people often just default to the smartest model and don't really try to reduce the costs when it's not being run at massive scale, just because the quality is normally so much more valuable than like $0.10 here or a dollar there. Awesome.

Jeff Kaminsky
Thank you.

Max Brodeur-Urbas
Appreciate it. You're muted then.

Ben Tossell
John. Sorry. John, do you want to ask your question?

John Reese
Sure. Where can I find out more about the exact capabilities of your advanced web search? I run into numerous problems with other web search subroutines or modules in that. Simple websites you would think like weather.com, wonderground, certain financial websites like Morningstar, rather than giving back actual data values, let's say temperature, wind speed for weather sites, it gives back JavaScript code which then goes to retrieve it. Now normal browsers, Chrome browsers will handle that and display it to the user. But where I found the problem is the AI actually gets handed back to it, the JavaScript code to get it and doesn't know what to do with that, never gets the actual value that we intended. How do I find out your exact capabilities and can you handle this particular little problem?

Max Brodeur-Urbas
Yeah, good question. So we make it super easy to just try out the product without even having to put in any credit card info or do anything beyond just signing up with your Google account. So I would, if I were you, go to use our website scraper. There's two different ways you can use it. One is the basic website scraping, which I think would cause this confusion with returning the JavaScript and not rendering it. But we have this setting for advanced website scraping which is using a much more advanced service under the hood that should be getting the content much more reliably. So I would say make a flow, paste in whatever website you want here and see what it's outputting and you'll get a sense immediately of every website we can handle.

Jeremy M Phillips (hehim)
Okay?

Max Brodeur-Urbas
Yeah.

Max Brodeur-Urbas
That would be my recommendation.

Max Brodeur-Urbas
Thank you.

Ben Tossell
And Kevin, you have your hand raised.

Kevin A
Hi Max, thank you so much for running this. Thank you so much Ben, also for hosting this. I'm from Indonesia where I am we have a brain drain problem where a lot of people are not really smart with all this stuff, even though it's the future. So I'm wondering if you guys have a consultant or somebody or an expert that can actually teach some of our guys or actually implement what we need and what type of technical levels that are needed for someone to do it, like a graduate student or like a fresh grad or, I don't know what type of qualification someone would actually need to be able to do this.

Max Brodeur-Urbas
Yeah, so we just released, we started our Gumloop Experts program last week, which is basically I just called out to our discord server where some people are spending upwards of 5 hours a day on the platform, automating workflows for their business or just for themselves. And I had them fill out a type form where they explained their most complicated workflow and recorded a walkthrough of it, and then have been qualifying power users as gum loop experts, people who would be willing to do contract based work, teach others how to use gum loop and build flows for a fee. So we do have that expert network and what I've been doing is just receiving requests via email and then I pair you with a gum loop expert. Not the most scalable system, but at the moment it's been working.

Max Brodeur-Urbas
So they would be able to help if you needed. And generally I would say the platform as a whole is approachable if you spend like 30 minutes tinkering with it. And it definitely does help to be semi technical or a pattern we often see is someone who took a python course in college or ten years ago and they know what a for loop is they know how powerful scalable systems are, but they just don't know how to actually execute anymore or write code. So gum loop is a nice alternative for them. Right.

Kevin A
Thank you so much.

Ben Tossell
Awesome. So is that people just emailing maxenthut dot de v or is it@gumloop.com dot gov.com dot?

Max Brodeur-Urbas
You can do both.

Ben Tossell
Yeah. And then yeah, you'll pair people up. And I don't know how. If you've seen people try and do this where they feed the documentation you have into claude or chat GPT and then they do that screenshot of like, hey, I'm at this point now and this thing's not working, how can I do it? Have you seen anyone get good results with using AI to debug their current thing? I know, I've done it with zapier and things like that before.

Max Brodeur-Urbas
You mean if any gum loop users have used AI on their own to debug their gum loop flows? Yeah.

Ben Tossell
So if I was trying to set up a flow and I was getting stuck and it was taking me longer than 30 minutes to get anywhere, just putting your docs into Claude and then sharing a screenshot and saying, hey, I'm trying to do this thing but it's not working, what should I do?

Max Brodeur-Urbas
I think, I mean, maybe someone has done that and if it works, I didn't hear about it because they never reached out. So hopefully that did happen. But I think most people just default to the discord. We have a technical support channel and we answer pretty quickly there. That's normally where you default to if you're totally stuck.

Ben Tossell
Okay, cool. Yeah, and I think, yeah, this expert program, I think it's probably something that people want to go to and. Yeah, go ahead, John.

John Reese
Max, what's your email again?

Max Brodeur-Urbas
Maxumloop.com. So just my name, thank you.

Max Brodeur-Urbas
Okay.

Max Brodeur-Urbas
Yeah, there's an m I think you wrote.

John Reese
Yeah, I mistyped it.

Ben Tossell
Yeah.

Max Brodeur-Urbas
Okay, cool. We also released a gumloop guide program, which is something we haven't publicly put up. I was going to write a blog post about it today, but if you go into your settings and then go to Gumloop guide, you can apply to be a gumloop guide, which is, I mean, it's annoying to say too fast, but it is a way to show people how cool Gov Loop is and generate revenue for yourself based on the affiliate link you provide them. So they get a 20% discount and you get 20% of the revenue for their subscription for twelve months after they sign up so I was doing the math yesterday. If you signed up 30 people for the pro plan, you would be generating like one eight K USD every month in affiliate revenue.

Max Brodeur-Urbas
But the reason we built this program is because the hardest part about gum Loop is just showing people what to do with it and all the different use cases that can be tackled with it. So if you wanted to pick a niche like you wanted to build SEO automations and show them to your SEO community, that's a good way to do it and be compensated for all the content you're generating.

Ben Tossell
Awesome. There was a point where in Zapier's life where they were trying to enable people to build a zap and essentially sell that zap, you could see how people could build very specific flows for different niches and effectively charge on it, but doing it the way of that affiliate.

Max Brodeur-Urbas
Yeah, we're going to do that as well. What we're going to add is an author name at the bottom of these flows and the ability for you to publish your flow, write the copy associated with it, and then if you wanted to associate the cost so that people can buy your flow for $5 or $10 if it's intricate enough.

Ben Tossell
Interesting. And yeah, Dave just asked this to be the last question. Looking at the process incoming invoices flow in your hub, is the AI reliable enough to process this without hallucinating concern about entering incorrect tax returns?

Max Brodeur-Urbas
I think the task is constrained enough that it is very reliable. I've almost never seen it hallucinate unless I'm putting in a handwritten receipt that is obviously difficult to read for a human, it's generally every receipt that I process. I think I've done a couple hundred while building that flow and then the customer who's using it has done a few thousand. Reliability has never been an issue. The only time it might be an issue is if you downgrade to a really cheap model like a GPT 4.0 mini. It might confuse a one for an I, but when you're looking at numbers, it's not that confusing. It's not like reading a password or something that's pretty reliable.

Ben Tossell
Great. Well, yeah, Max, really thankful for your time and showing us all this. I think there's a few people here who are the wanting to do a bunch of stuff here and learn more and try and build a bunch of flows and yeah, please share them with the Ben sports community and with Max. I'm sure he'd love to see what people are building. I think you've probably got a few takers for your experts program as well, so expect a few emails there. But, yeah, I really appreciate this. And we'll get the recording, transcript and everything else up on the site as soon as possible. So be up there tomorrow by tomorrow to rewatch and dive into anything but, yeah. Cheers for your time, Max.

Max Brodeur-Urbas
Sounds good. Nice to meet you all. Fair enough.

Ben Tossell
Thanks, everyone.

Max Brodeur-Urbas
Bye. The recording.

‍