# Studio Update #05: Leaving Make.com for n8n, Finetuning Small Open Source Models, ITOps automations

## Метаданные

- **Канал:** n8n
- **YouTube:** https://www.youtube.com/watch?v=68ULIfj4Pjw
- **Дата:** 10.01.2025
- **Длительность:** 21:37
- **Просмотры:** 6,402
- **Источник:** https://ekstraktznaniy.ru/video/15501

## Описание

In this episode, Max and Angel dive into new projects for 2025,  and a chat with an automation expert who switched from Make.com to n8n.

Connect with Max on LinkedIn: https://www.linkedin.com/in/maxtkacz/

Here's what's inside:
🤖 AI Agents, Part 3 - Get a sneak peek at Max’s most in-depth tutorial yet—prompt engineering for agentic systems.

⚙️ Fine-Tuning - From 30 to 100 training examples, can small open-source LLMs be made to perform surprisingly well? 

🌵 Angel’s Arizona Update - MITRE ATT&CK powered agents, and templates for ITOps and SecOps

🔀 Moving from Make.com - Automation consultant Anthony Lee opens up about what he loves in n8n, what he misses from Make, and how to bridge those gaps.

Chapters
00:00 - Intro
00:13 - Max's Project Updates
04:47 - Angel's Project Updates
09:00 - Anthony's switch from Make to n8n
20:33 - Wrap up


🔗 Links and Resources:
Sign up at https://n8n.io and get 50% off for 12 months with coupon code MAX50 (apply after your free trial!)
https://commun

## Транскрипт

### Intro []

hey it's your favorite flowr Max and it's literally snowing sideways right now let's kick off the first episode of 2025 with some studio project updates from myself and

### Max's Project Updates [0:13]

Angel on my side I'm working on part three of the building AI agents tutorial Series this third part is on prompt engineering and specifically prompt engineering for AI agents so there's definitely some different guidance when you're working with AI agents building your prompts versus just LM so I organized it into three different sections basically the system layer the input layer and the action layer these are all the different places where prompting can occur it doesn't cover everything what I really tried to do I was doing hours of research over the break is trying to figure out like what's most important for most use cases and what's information that will still be useful in 6 to 12 months skills that you could take throughout your building 2025 and through your millionaire years into 2026 and Beyond I aim for that to come out next week but it is a mey tutorial it's probably the most in-depth thing I've done to date yeah really guting to have that out next week so we can pause the building AI agents tutorial series get some feedback on it and get into some other projects on that note one thing I really want to do this quarter is help some real teams solve some real problems and I was thinking which teams I could help it hit me there's a bunch of teams at netn with real problems and I'm going to help nadn support team this quarter automate some of the pain points that they have using AI in aut so we had a quick chat with the support team and what we're going to do is have a listen to their problems and together basically do like a bit of a discovery on those problems triage those issues and basically come up with a list of things that we could do pick one of those scope that down build it ship it iterate on it and show you how a real department and a real company with real user data so there's going to be some stuff bled out but basically how we solve oural problems and I'll upload those to the template Library really looking forward to get going on that and I'm meeting with NN support team next week on that so we'll probably have a little sneaky update next week and then another big theme for q1 perhaps even throughout 2025 for me is going to be exploring fine-tuning specifically the whole world of taking small open source models taking equally small training data sets like we're talking 30 50 100 examples and then getting like non-trivial Improvement on those small models so a great example is shedy copti um this is one of the folks I follow in LinkedIn you know you see all those memes that say like linkedin's dead linkedin's garbage there's no value on there it's only shelling he is like an antidote to that and I was Fanboy liking every single one of his posts over the break but basically he did exactly that he took 130 synthetic examples so that means examples generated by an AI but like a frontier Ai and then uh basically took that training data trained his I think it was a 3B llama model trained that for 3 minutes on those examples and got like a 40% increase uh or something like that on some eval Benchmark so of course benchmarks not the same as the real task and whatnot but I want to explore that because at a high level from my research so far you're basically going to get the data be it's synthetic or be it from your different systems and it ends pretty good at conf fetching data from systems and transforming it in the right format you then got to use some libraries like pie torch and stuff to perhaps run that training now again NN can do that if run locally it can run shell script so we can kind of automate that and get all the different arguments and stuff for all those calls complicated calls a bunch of stuff that most people don't know about and get some decent presets for those and pretty eject those then we need to interact with some apis to upload it to somewhere we can deploy it lots of cloud providers now allow to upload you know model files and that kind of thing so I got to figure out all the details the model file names the process what all these presets are probably going to try and work with some experts on this kind of stuff cuz my job is to basically translate that stuff roll it up and make it understandable but in any case we're going to be looking at fine tuning and exploring can NN be used to orchestrate your whole fine-tuning pipeline I don't know if I use enough industry terms there I think I did um and if so we can then basically in the community you can take that General template and make your own connectors um and what I Envision you could do with something like that if I'm successful is for example have a new model each week that goes out for each of your customer support people in their voice now perhaps that is indeed Overkill but I want you to be able to make that decision and I want it to be in your power to be able to fine tune a model that takes 4 GB of RAM to run and $3 to train I want you to be able to make the those calls now some people have expressed um fear of me getting run over so look I'm crossing I'm checking good all right what's next uhhuh so since I'm dealing with pretty dzian weather over here let's head over

### Angel's Project Updates [4:47]

to Arizona and check in with Angel Menendez hey Max hope you're doing well so first of all I wanted to share some progress on my miter attack project as your viewers know this is something I've been working on for a month or two now unfortunately I've been running into delay after delay but thankfully I've finally started making some progress let's talk a little bit about what we've done one of the biggest issues I had is that the miter attack Json file that I was working with was just too big it was about 40 megabytes even nadn had trouble parsing such a giant document so I finally made progress by utilizing chat GPT to generate a python script that enabled me to just extract the parts of the document that I cared about reducing it from about 40 megabytes in size down to about three so that size difference made it much easier to interact with an n8n parse it and then from there I was able to send it into an n8n workflow and map it into quadrant one of our partners in terms of vector store so from there what I was able to do essentially is extract the relevant parts of miter that allowed us to then map those into a vector database so I'm really excited to share what that database looks like there's some really cool visualizations within quadrant that allow you to see the database in a 2d space as little points and those points represent how distant the different keywords or objects are to one another which is really cool to see in a miter kind of framework so from there what we're going to do next as part of hopefully our next Studio update I'm going to be then connecting that Vector store to an I agent as we had talked last time miter attack is like a catalog of different attacks that a an hacker might use to penetrate an Enterprise organization's defenses by mapping these different forms of attacks as catalog into a vector store what we can do is provide an AI agent with all of those attacks in the form of a vector store for it to query every time a question comes in related to a cyber security incident so from there you're able to then get information such as what kind of attack it is and different ways to mitigate that attack essentially allowing you to automate a lot of the ticketing process that you might have internally to help your it or cyber security teams essentially mitigate these attacks much faster because you know what you need to do to take care of that attack you don't have to do that level of research in advance so something really cool something I'm very excited to share so one of the other things that I wanted to share is on January 22nd we're going to be hosting a webinar partnering up with Voice flow a leader in AI customer support agent technology It's a Wonderful software that allows you to take in voice calls real time and connect that to nadn to essentially take care of some of the external API connections but process a lot of the voice AI portions on their platform so I'm really excited to partner up with them we're going to be demoing some use cases we're going to be showing how those use cases are run and hopefully our viewers will get some benefit out of that so another thing I wanted to share is that I've been releasing more workflows for nadn for our users to download for free things mostly aimed at it operations or cyber security customers so things like analyzing email headers within nadn suspicious email analysis using chat gbt Vision things like analyzing suspicious email content using chat GPT itself and even some new workflows working with service now so if you're using service now as part of your internal flows I think you're going to want to check out some of these workflows so I hope that gets you all excited for what's coming up I'm really excited on my end this is going to be a fantastic year I can't wait to show you all the things we've been working on internally here and I wanted to do this video outside but it is freezing here in the desert it is 32 American degrees here so it is ridiculously cold for us I am definitely not going outside right now hopefully our next video I can have some nice footage of the lovely Arizona desert so back to you Max hope you had a great holiday hi thanks for that one Angel so

### Anthony's switch from Make to n8n [9:00]

for this next bit I had the privilege of talking with Anthony Lee who was an automation consultant that was using make. com well that is until he found NN now there's a lot of different ways that one could record a how I switch from a competitive video and there's also plenty of marketing out in the world right now I don't think we need more so what I did instead is I got on the phone with him I said forget how you think this kind of content should be done and just give me like a r candid what you like what you didn't what sucked um and we'll edit it down and keep the bits that make in not look perfect too and show it to the people and have them all make their own opinion so I don't know Louis you might not like this but he says a lot of nice things about naden and one or two things that he misses about make which we should probably make better chy check this one out hey Anthony how's it going it is fantastic how are you doing really well in this bright and early Monday thanks so much for joining the call I appreciate it it's a pleasure to be here thanks for having me fantastic you were one of these people Anthony where you know I was seeing some of your posts on LinkedIn and on LinkedIn there can be a lot of people Shilling their thing makes sense but you one of these people was like wow that's really interesting I'm clicking that show more button and so I wanted to bring you on the show to hear a bit more about your automation experience and in particular I heard that you've been making a bit of a switch from make to n ATN but before we get into all that would you mind introducing yourself everyone absolutely all right so my name is Anthony Lee my background is versatile I came from the e-commerce space and then moved into automation which I was actually automating things before AI became generally commercially available so as soon as that happened it was a very natural and that's now what I do build these AI powered robots for lack of a better term and are you focused on Ecom for your uh AI automations I do a lot of work in Ecom because that's where my network is but definitely venture out of there cuz as you know automation touches everything everybody has a finance department that does way too much copy and pasting everybody has you know content creation desires and every single industry can use these things so I haven't focused makes a lot of sense and so I heard that you were a big user of make. com as part of your animation work and you're using nn more and more now could you walk me through some of the reasons why you started looking for other tools Beyond make absolutely so I actually have a very long history with make I was using it back when it was ingrat and at the time that was honestly it was a budget decision it was just a lot less expensive than zapier so that's where I started and the platform did grow and there was a bunch of desirable things about it and it got better so I stuck with it but ultimately what happened was I started running into a lot of limitations okay so little things right for starters uh now that we have access to AI pretty much everybody can run code functions right functions are just little Snippets you don't have to write a whole program you just needed to do this one thing and it's so simple um but make didn't at the time and now it kind of does but still not very simple not very easy to implement uh run code and that was the biggest wall I kept running into like hey I have to do this thing that's not a native function and that makes sense but I can't even code it in there um and then another one which was really big for me was um you can't loop it's all linear uh so I started reading up on na10 and to be perfectly honest I wanted to move over a long time ago okay but the reason I didn't was because a lot of the documentation for na10 is intimidating it looks complicated you can do the cloud version but really everybody does it local and then you have to you know open up an environment window and you have to put in this even though it's it's literally copy and paste one line and then click enter But at the time intimidating right okay I got you yeah so you know and then and if you read any of the forums they talk about how oh you really want to do this in Docker and they don't talk as much about mpm which is arguably simpler but both of them are really simple so it took me a while but finally I was like no I'm going to do this because I really have needs for being able to implement code functions and being able to Loop those are the two biggest ones then I started playing around and it was just man it's so powerful cuz I can do everything that I need to do even if there's a node in na10 which and a lot of nodes that do a lot of great things but even if there's an aren't nodes for something I need to do just for I to you know have Claude write a JavaScript function and it does it parses everything puts it where it needs to go runs it as many times it needs to run and I haven't been able to look back since what was like that first aha moment where you like okay you got over the maybe scary looking docks and you got in there what was that first moment you're like wow this is powerful this is looking like something I want to explore more I think the biggest aha moment for me was I was working on this huge workflow in mate to build content using Ai and the problem was is it was very rigid because again it's all linear so it's like all right the outline always has to be this long it separated in these sections and I didn't like that rigidity so one of the first things I worked on doing inside of na10 was I just want to see how I would do this you know real simple Loop function and that way the AI could write whatever was natural and it's weird to use that word with AI but whatever like however you know whatever the length of the outline needs to be and then it could generate each piece and um when I did that and it worked and was like this particular topic happened to generate 12 sections it was huge article I was like oh my gosh like I can literally have this system combined with AI create all the content that I want and that was like the beginning and then from there I've done a million other things gotcha how many workflows around do you have like running right now well so my own environments I only have about I don't know 25 or 30 and only five of them probably at any given time are in use so I have one client I think we run like 15 different workflows cuz I built a entire system for them and all of the flows are big but you don't want them to get too big so it's like all right this one triggers this one so it's huge all total probably about 35 running this is a meaty chunk of flows how long have you been in the N andn Community around oh gosh it's probably only about six or seven maybe eight months now okay wow that's a good velocity there I haven't used me that much myself one thing I do hear from people is sometimes on the sort of cost comparison between running the workflows can I get handle from you on that like did you notice any sort of costing differences switching to Ned in for Mi I mean yeah I run it local I don't have the monthly fee but also so what was interesting was na10 was the reason why I downloaded no Js on my computer and started using npm and because it was so simple right like literally copy this line of code paste it here click enter I was like I can do this with other stuff so because of na10 I started using other open source softwares that integrate with na10 so I don't pay for you know NOCO DB mongod DB there's a bunch and it's just like it's all on my computer now app Smith like different things so that has significantly reduced the cost of running to be fair it is easier when you don't get to sit at the computer work in the cloud so my clients all have Cloud instances those Community nodes can come in there's some people building you've G you've gone into the community nodes tell me about that experience what if oh by the way so for anyone watching Community nodes is User submitted nodes that uh the community submitted that you can download and run I'm curious Anthony what are those nodes are you're like oh my God I needed that pop it in got your job done so the biggest one that I use all the time is global constants right because you need to be on Enterprise if you want to have Global varials but constant ows me because I have a handful of and blah blah blah it's pretty long list but I need them for everything so Global cons has been the biggest one and then there's a couple of little other ones that just like transform text and but I'm finding new ones all the time like great now I don't have to write a function for that so so Anthony firstly thanks for all the amazing feedback obviously as you can imagine that someone who worked at Ed for as long as I have it's a pleasure to hear all this nice stuff but I'm curious like if you had one magic wish to make nadn the perfect automation tool big effort small effort doesn't matter if you could wave that oneand what would you change oh man that's actually a really good question uh I mean at this point I don't know if I'm still in a honeymoon phase with it or not but uh it's awesome I would say okay so one of the challenges I've run into is extraordinarily not userfriendly for a non- coder to connect meta products right so for example in make. com if you want to connect your Instagram for business it's the you know single sign on window you go in and you're in but I just recently went through this incredibly arduous process with na10 where I had to go and I had to create an app do like a ton of very technical stuff I would change that like that would be great if the Social Media stuff was just a little bit easier uh okay so this is like dealing with the off layer and then having to create like a client ID and see create an app on your own or makes sense so yeah thank you for that feedback Anthony I'm going to make sure that it gets to the product team at the end so I app so Anthony if people want to follow along your automation Journey or see the kinds of things you're building where can they do that I definitely would Point people to LinkedIn that's where all of my social media focuses right now you could see everything that I'm doing because I pretty much share anything that I build on there 2025 I actually do plan on making na10 templates available through my website AI plus automation. com which all of those links and everything for AI plus automation is also on my LinkedIn cool and then there'll be a link wherever you're watching this is going to be linked somewhere to Anthony's stuff somewhere here Anthony really appreciate your time sharing your experience switching from make to n at end I think it carries a lot more weight when it comes from someone that actually did it versus someone who like works at n at end and I really appreciate also the constructive feedback on things that were easier to do and make because again I want to make sure that in this year we make that also easier for all Val Jesus good luck in all the AI automation building efforts this year I think there so much cool new tech came out even just over Christmas I'm still catching up but I'm really excited to build this year definitely and once you've uploaded some of those cool templates you're talking about I'd love to have you back on the show Absolutely I'd love to Fantastic well have a great start to week Anthony absolutely you too thanks bye cheers and then there's a kind of awkward pause so I'm gonna Anthony thanks so much for your time again and I've already gotten that feedback off to our Cloud team and we're going to see what we can do about making or easier on the few apps that we haven't covered yet with that sexy click

### Wrap up [20:33]

to connect experience all right everyone that's all for this week what I'm doing the next two weeks is front loading a lot of interviews so we got some nice uh wiggle room there I'm finding a video editor so I have more time to build so hopefully next week you'll also see more stuff happening and I'm working on the building AI agents part three which is prompt engineering for AI agents or rather for agentic systems so keep your ears peeled for that one ears peeled keep your eyes peel for that one and what's my sign off in any case I'm Max your happy resident FL grma and I can't feel my fingers anymore so I'm going to go but have an amazing week and we'll catch you next week for another Studio update happy flamming everyone by