Introduction

So the title of this presentation is Beyond the Buzz, where LLM meets life, life meets LLM, what are the LLM's strengths and weaknesses.

And this is in a nutshell a representation of LLM, which is super smart, trained on billion, billion, data points, parameters, text data, trained already on your conversations, on your source code.

You may not know it, but LLM may have been trained on the data that you sent to the Internet. But in a nutshell, it's super smart, but it's also clueless.

LLM and Assistance in AI

And the latest trend of all LLMs of all of the AI companies, being it OpenAI, Dust, Mistral, DeepSeq, et cetera, doesn't matter. They all try to promote the trend called assistance.

We may see that as, for example, deep research. We can see it as the background task.

So effectively, what it means in a nutshell, it's an LLM plugged and connected with something else, something that makes it not just spit out text, very precise text, very good text with fast speed.

So the amount of text with the same velocity can probably never ever be achieved by a human being. But it doesn't really know what to do next.

Workflows and Automation

So on the other hand, there is a concept called workflows, right? You may have seen that in different formats, like Make, NA10, probably if you have software engineering experience, you already coded the workflow for your company, right?

So you encoded some business process into the computer so the computer can accept the data, process the data, output the data. In a nutshell,

Yeah, it could be as simple as pulling in client details into Excel spreadsheet and making a report. This is an example of a very simple workflow, but depending on business complexities, legal regulations, other sort of constraints, workflows can go as complex as that.

Sometimes it's the downside of the business process, sometimes it's the downside of the tool that has its own limitations. But usually that's what you lately see on LinkedIn as advanced workflow that you can share, promote, use in your company. 1And this is something that actually saves you time, right?

Integrating LLM with Workflows

But how does those two play together?

So how can LLM understand something like that?

How you can execute things from the chat, with your voice, just freestyling, thinking, and get things done.

So probably LLMs are not yet there to understand this complexity, but I'll show you how, for example, deep research is implemented in very simple terms during the practical example, and how you can make more complex business processes understandable to LLMs and how you can access quite complex things from the chat interface.

The Concept of Simplification

So here is a term I borrowed from software engineering. I promise it's the only software engineering term I will show to you.

It sounds very simple. Keep it stupid simple.

So how I'd like to compare LLMs and generally workflows that you try to put and try to optimize something in your company for your private life, for your side project, for your business. I think about those things, about tasks and about task orchestration as about team of interns.

They just came to you straight from high school or straight from university. They're clueless and probably they wouldn't comprehend something like this.

Breaking Tasks into Simple Steps

So 1for example, we today will break down the concept of curating newsletters. So curation of newsletters is a very simple process.

Lots of people in marketing, in just private blogs, need to curate content. You run out of power, you cannot write always the content, but you need to always look for sources of inspiration, you need to share links, you need to exchange links to build up your presence on the internet. So it's a very valid business use case, and later on we can see how we can reframe it to cover even wider space of business problems.

So I imagine for newsletter curation, I would need a team of interns. I would delegate one intern to look for articles on the web. Very, very simple.

One intern like you. You go and look for the topics on the internet. Full stop.

He knows how to do that. He may produce some weird results from time to time, but we are here to evaluate. We are the orchestrator.

The second intern would go through the links provided by first intern, provided by yourself as an orchestrator, and just download articles to build up your archive. Break down, we are trying to break down tasks so it's understandable and repeatable by a single person, so they do not run off the power, they don't run into the cognitive constraints.

LLMs as Intern-Level Task Executors

The same works for LLMs, right? LLMs have context windows. LLMs are very good for repeatable, simple workloads.

The third, Internet would save information for you on Google Drive. Fourth, we'll look for, let's say, if there are some more complex queries, you need to to be able to send a team of interns for a week and forget about them, and then check results.

So that would be the task of the fourth intern, to just send another team and forget. Next on, it's more like a task for an orchestrator, so I'm describing work for myself. I'm going into some sort of control panel and checking status of the intern team.

So literally it would be an equivalent of having a meeting. Okay, give me the list, give me the articles, give me the content. Okay, now I need to produce the newsletter.

And well, the last thing is very simple. We have a database of content with links, with description. Let's produce the newsletter.

So that's how I see the workflow of curated newsletter list in practice. And

Instead of one large workflow, like you've seen on previous screens, I use very, very simple concepts. Look on the internet. Download the article.

Save the article into Google Sheet. Very, very simple. I keep it constrained.

I keep it understandable for a one intern with no experience so I can literally give him a task in one minute and in five minutes I can verify the results.

practically accessed it.

Demo: Practical Implementation

So finally the time for the demo. Let me show you how I configured it for myself and what I used.

I used literally those three workloads in Dust and I used as an interface because I need to orchestrate the thing. I used Dust TT, so it's a tool that helps you to build the content, sorry, to connect to the external tools and basically instead of the clueless robot, You get a robot with the access to your information, the robot with the access to your Google Drive, the robot with the access to the Google search.

So give me a second. What I'm probably going to do, I'm going to change the settings of the sharing so I can see what you see. Give me a sec. For now, any questions while I'm configuring it?

I probably got only one. When you mentioned deep research, what application did you mean? Yes. So deep research, for example, if you go into deep research mode in Gemini or in Google, deep research is a more advanced step-by-step agentic system. So it's LLM with plugged tools. And I'm going to elaborate right now the concept of tools.

So it has access to web search, it has access to downloading the article, and it has access to a few other things that produce big, large, fancy reports. Of course, it has access to more processing steps. So LLM can do a few things. LLM can reason, so it can create a plan for the execution.

LLM can have a few iterations, so it can download the article, it can summarize the article, and next step, if, for example, it sees that article's summarization results do not align with the initial prompt, it will repeat the step with the different parameters. 1So deep research in a nutshell is a tool-fueled engine that runs step-by-step more complex tasks that a simple completion that you usually get access to from the LLM interface will not be able to handle. So that's in a nutshell what it is.

What is the accuracy of the research in the deep search, based on your projects, on your experience? On my experience? I will address this question as the first one after the demo, okay? Because it will be aligned with what I show right now, because it very much depends.

Okay, so what I ended up creating... we have all of these scenarios. So we have the basic search tool in Make.

Who has ever touched Make? Raise hands. So yes, NA10 or any other automation tool? So half of the audience. Effectively, I'll give you words about that.

It's a tool that allows you to access your Google Drive. It's a tool that allows you to orchestrate further actions. So I have troubles connecting to the internet. Sorry for technical issues. Yes. okay let's let's check it again okay

Yeah, I'll connect to my phone, just in case. because I need Internet to do the demo, actually. Unfortunately. I actually had an idea to record it on the video, but then what wrong can go with Internet in Warsaw, right?

Okay, it looks good right now. Yes, we are there. I hope I haven't lost your attention. Stay with me, stay with me. Not yet, good.

Setting Up Tools and Workflows

So what I ended up creating is a few very simple tools, as I showed to you. So based on the simple search tool, I'll explain to you what Make is as well, so you're more aware and you're more equipped. So in this case, the tool is super simple. Again, keep it stupid simple.

It connects to SERP API. It inserts the query. SERP API is sort of abstraction for Google. Why I use SERP API, not Google, it's faster, more higher rate limit. But you can use whatever you want. It doesn't really matter for this purpose.

So I can test it. For example, query articles about AI. And as the result, we are seeing some articles in very unnatural format. But it works.

And just a hint for the future, what I'm doing so that LLM can use that. straight out of the box. I'm explaining, do you see it well? Yes. I'm explaining what this particular workflow expects.

So it expects one parameter called query, and I describe it as search query. And what it outputs is search results, search results from Google in JSON format. In JSON format because I decided so. It doesn't really matter for LLM in which format, but I like to keep it structured. So what does it mean?

On a very simple example. Of course, Dust is another tool. It's, again, an agent orchestration engine, so you're able to create your own agents in there and plug in your tools. So I created a very, very small agent for demonstration of one tool, search. I called it web search.

Just search agent. So I can tag multiple agents in this particular interface. It's very simple. And I can ask it to find me five latest EV models.

And what will it do? Through the MCP configuration, Model Context Protocol. Model Context Protocol is an abstraction between Make and LLM. In this case, Sam LLM by Dust. I don't even know which one, to be honest. It doesn't really matter because Model Context Protocol makes it simpler.

It uses the basic search tool, and it will format the results. Let's give it some time. But in general, in a nutshell, I'll show it to you. I'll prove it to you.

It calls your search tool. And it doesn't stop there. So you can create any kind of tool. For example, if you want to access from the chat interface, your spreadsheets, your Google Drive, your big data database.

You want to write scripts to write data visualizations. It's all now possible, but through some abstractions. So, yeah. Pretty accurate.

And I didn't do anything else. Like, that's it. So, to prove to you that it actually ran it, history, it ran just 30 seconds ago, and you can see, because Mix saves all of your workflow executions, search results, Yeah, here you go. Voila.

Executing and Verifying Tasks

So very simple example how you can make a simple search query.

At this point, any questions? What is a V model?

Electric vehicle, like Tesla, like, yeah. Yeah, it doesn't matter. You can put anything there.

Yes? I have a question about the red flag in the dust. Yes. Yes, yes, good question.

I'll show how to do right now. So over here in Make, I showed to you very intentionally when you edit things. Let's get back to basic search edit.

You remember how I described input and output, right? Also what I changed, I changed the shadow in settings so this workflow can be executed on demand. For engineers, it's an alternative of creating your own API.

So if you're not an engineer and you create this workflow, you can brag to your friends that you created your own API with no code tools. Within five minutes. Literally, this workflow takes you five minutes to create.

In Dust, okay, so not everything in Make. In Make, you also need to create your MCP key. I believe it's available for core plan for sure.

Not sure about free plan, but core plan for Make costs you like $10 per month, so 40 zloty. So you go to keys.

Oh no, it's not this one. no, sorry, you have to go to profile, you have to go to API and MCP access, and you have to, it's already my MCP that I created for the purpose of this presentation, but you can create a separate one, it's again very easy, it doesn't require any configuration, but it requires the description for every workflow and configuration so it can be executed on demand.

And that's it. So next step for Dust, This is not the talk about DAST platform on its own. I don't have enough time because it's quite powerful platform.

I believe it's a French company. Very clean, neat interface. I use it quite a lot.

But in general, you can connect your MCPs. It has its own built-in tools, but Make creates MCP server. And in Dust, you can connect and use MCP client.

And in a nutshell, it gives you access to all of the tools that you create in Make. If they're well described, then LLMs can access them. So it's Make MCP that I created for this purpose.

So I connect. to server URL I get from make settings.

I just use default settings and it still works. And what's critical here and what you have to verify is all of the tools that you receive from your workflows. It must be all of the tools you set on demand.

So over here you see all of the set of small tools that I created. And let's just briefly walk through them.

So web page retriever. Only purpose of this tool is to download a web page. That's it.

Clear, simple. Intern level task.

Basic search. Enter search query into engine. Basic task. Receive results.

Create background research. And now I'll walk you through that. So more complex case.

I'll show to you in one second what would happen if you don't do some tasks in the background. You'll simply run off the context window and you'll simply run out of a session time. Because let's say in the example for electric vehicles, you've seen that it took the LLM like around a minute to process it and it was limited by 10 results.

So with the increase of the content, you need to give LLM more time and you need to give all of the other workflows that are connected for the purpose more time to get done. So you need to sometimes send a team of interns for a week, lock them in a room and check when they're back.

Advanced Applications with LLM

So what I ended up doing, how you as normal people without access to fancy databases, background running tools, can create something like that using Google Spreadsheets, Make, and just Google Drive. Very, very easy.

Everyone pretty much has access to those tools, or to Microsoft. Depends on what camp you're in. Doesn't really matter.

Make can be connected to any of those. So I created a tool called Create Background Research.

Two steps. So what it effectively does, it manages this table. And it's a very simple table.

So ID, some random number, or random text, topic, so your search query, date, for tracking purposes, and the status. Either it's 0 or 1, or for some reason, probably there is a bug in the system, it should be 1, a date. So...

Newsletter Agent and Background Tasks

I created another next level agent, a little bit more complex than search agent, so let's get back to that. I called it a newsletter agent.

So over here I've done some configuration changes. I'll walk you through them. This is a system prompt.

Yeah, do not hallucinate, I explicitly asked. I explicitly asked LLM do not hallucinate and specifically use the output of LLM.

Why that is important? Because if I didn't put that, LLM, even if the tools failed, would start producing some jibber-jabber. Like, I don't know, something it was trained on, it just... LLMs in a nutshell are very good at predicting the next word of the sentence.

And they will create very coherent text. They'll create a text you would believe is true, which is not. In the reality, I need to explicitly ask it to only use the output from the tools.

which are factually correct because it's Google. So what I connected it to, I connected it to MakeMCP, so it has access to all of the tools.

I connected it to ReasoningModal, so it's Google Gemini 2.5 Pro, I believe. File generation, just in case, not used there actually.

And include data. And include data... effectively it's connected to my Google Drive.

So my Google Drive, this one. It has already some articles, it has this spreadsheet, and I explicitly gave access to the agent to use this information.

Creating Background Workflows

So let's start our first background workflow and walk through that. So that will be this workflow. And from Dust,

I'm asking newsletter curator to create a background search job to get articles about any topic? Cucumber. Cucumbers. Let's make it more complex. Growing cucumbers.

This is unnecessary. And let's see what happens. So it's thinking. Let's give it some time. It generates steps.

And you see I marked this as high risk task because it actually changes the spreadsheet. So I need to explicitly allow human in the loop. I'm here for something orchestrating. I need to confirm that it can change the spreadsheet.

Growing cucumbers. Yeah, it's done that, right? So the agents finished, and that's it.

So what happens next? Next on, I created the only probably workflow in Make that is not connected to MCP very intentionally. It's called Index Research.

And what it actually does, every two minutes, It launches and it checks changes in the spreadsheet. It takes the topic from the spreadsheet.

Yeah, it's seen. It walks through all of the search results and I believe I set the limit of 100 here. It gets the page and it saves the content in the Google Doc. for every search query.

So of course, it will not save 100. Some pages will protect you and will protect themselves from the bots. And you will not simply get access to this new trend that OpenAI parses everything, creates loads on site. So lots of sites started blocking bots. And we are bots in this case. So we will get blocked.

But let's see how it dealt with the task. if it's dealt with the task, index research history, and it's running. So for this purpose, usually we're lazy orchestrators, we don't like to go into the nitty gritty details of spreadsheets, so I would not use it. But I can ask, get the status of background tasks.

Yes, it is possible. It's a very good use case. No, I did not implement it. But if we have time, so Jacek, please control me on time if I have enough. Yes, so you can do that.

Even if we don't get time today on the session, I will post the blog about this particular topic, and I'll add this use case, especially for you for the first question and request. So, yes.

in progress and a bunch of things that I tested it on as completed. So we need to wait for it a bit so it goes through all of the results. But effectively, what it's going to do, it's going to download articles into this directory.

So you see it's jumping, it's increasing its size. So we start learning more about cucumbers. But I'll show you the last step of my demo.

Generating Outputs for Newsletters

So I'll ask Newsletter Curator, for example, you know that some tasks are finished, right? So right now you need to get the output.

So you are asking to get and suggest articles with links, short descriptions and titles for a newsletter email about machine learning. I know that machine learning articles are in the database. I tested it before.

So let's give it some time to go through the Google Drive. And in the future, if you want to make it better, basically what we've right now done, we created micro deep research. So we create a background task.

And we created... Yeah, I see that. So we created a background task and we are getting results from the previous background tasks. And here are articles ready to be copied and pasted for your newsletter.

And this is it for this demo. Any questions?

Conclusion and Final Thoughts

Is there a notification system implemented so that system lets us know when that job is done and when we need to check it every five minutes? Well, I didn't do it for the demo, but you can do it as simple as go into index research and whenever the workflow finishes, you create a block here to email yourself. So you basically create Gmail integration or Slack integration, whatever you prefer, and it will get you notified.

So it's a very extendable system based on a few small steps that you can put into use within a few hours and you don't need a team of expensive engineers. You need Dust for 29 euros a month, you need Make for 10 dollars a month and I'm using paid search because personally I have a bunch of agents so I'm paying for quick search capabilities but you can use free search engines.

Next question please.

And if I wanted something to leave with you, make things simple. Don't go into overengineering complex workflows. Very few small tools with the integration, with the power of reasoning LLMs can get you very long way.

Thank you very much.