Launched EmailSpy on Product Hunt, Hey Siri AI Agent template [Update #09]

hey Sprint Squad it's Max and it's the end of day 26 of the AI Sprint it's time for some project updates and as always on the AI Sprint a lot has happened since the last update first up we launched email spy on product hunt yesterday once again I want to give a big shout out to Oscar for all the work he did on launch day and in preparation for that I'll be honest Email Spy wasn't some sort of viral success we didn't get tens of thousands of people using it but as of right now there's over 190 unique URLs that were ran through it and I got qualitative feedback from folks saying hey this was interesting this helped me understand the capabilities of this tool I might clone this and duplicate it for some serious scraping use cases I have for my job and if you look at the first AI Sprint video that's what this is all about it's about having people know wow this is feasible you know in a matter of hours or days I could build something useful or something that's enlightening so with that being said I think it was a great success now you can duplicate the project there was a little bit of a delay in getting the front end in a way where you can duplicate it but that's now published at GitHub and it's in the official Email Spy duplicate page so again the whole project can be duplicated there there's the Puppeteer part of the workflow that isn't made public yet I think what I'm going to do is take the email spy backend workflow and make a simplified version that just uses one method so that's just the f c approach that way folks can set up just a few credentials when they duplicate that template and check it out and then look at the or original template if they want to create their own multi- approach strategy which again could be a combination of AI methods or traditional scraping methods and then this week Oscar and I got on a call in preparation for the launch where he went through his backend workflow and explained how it works so later in the video I've also got a cut down of that where he walks us through the workflow if you have any more specific questions as you're looking at that or you're looking at the workflow yourself make sure to ask in the comments because I can ask Oscar before the Sprint is up we can get on another call and answer more specific questions and one thing that happened last week that didn't have a chance to edit down yet was the interview with ol our AI engineer later in this video you can get a cutd down of that where he talks through some of the things happening in the AI space right now including some new models that dropped recently so if you're interested in hearing about that make sure to keep watching the rest of the video and then once Email Spy launched and was underway I had a chance to publish the hyi template today you can download it off nend IO and that comes the Apple shortcuts companion workflow so there's an apple shortcuts workflow and NN workflow the Apple shortcut lets you do the hey Siri part and it sends along whatever you dictated to the end workflow and then in the nend workflow I've got a really simple AI agent that takes that text and replies there was a few details in the template like the exact format to send it back so that Siri says it so I handled all that for you but it's a great starter kit for any kind of hyeri like template that you'd like to run in inside NM so in that template I show also how you could pipe in real world context to the AI agent I do it with a really simple example where I basically in the system prompt I tell it the day where in the system message I use some very basic JavaScript using the now variable so that basically tells it what time it is right now and well actually let's just show you hey Siri AI agent here's what I all right what did I call it ah ask hey Siri ask agent what's the text what time is it right now message the requested webook post assistant is not registered H okay so it's not working cuz I didn't activate my workflow for anyone that's a power of NN that'll probably be particularly funny to you the workf isn't activated so that's why the demo is not working however it did show that NN is responding to Siri because anyone that knows uh what was being dictated by series that's an endent error but in any case I digress tldr is it's a really basic workflow template so it doesn't do anything particularly fancy other than having whatever you say to Siri go into the AI agent and then whatever the AI agent outputs going back to Siri and has some light prompting so that Siri is concise doesn't output things like new line characters so really it's ready for you to go in there and add context so if kinds of things you could do for example is slap a notion node before the Siri 1 perhaps pull some details and pipe that into your system message then adding context you could obviously add tools to the AI agent as well it could search Wikipedia for example or your own CRM or something so an AI agent something like piri and then you ask about a certain contact and it goes searches that and returns that should be pretty simple for you to knock out now that all the basics are there since it's just a basic template I was talking about this here use case with Devon K because he's one of our awesome community members that I didn't have a chance to formally build something with during the Sprint so Devon your challenge if you choose to accept it is take my hey Siri template and use that as inspiration to ship a really cool Siri template you're working with customers every day on these kinds of use cases so I'm sure you've got plenty of inspiration for that this would be a great time to mention that we're going to be announcing a contest at the end of the AI Sprint where you can submit your AI workflo the first prize is going to get A1 gift card we use the service we can pick basically any kind of gift card you like but then 12 months of end Cloud Pro for free so it did the math that's like € 720 worth of work for Automation and we didn't talk about this on the team but if you win we can jump on a call and talk about your work for us of course in the next video I have some more formal updates about that but start building and I haven't figured out exactly the judging criteria but it's probably just going to be what impresses me and I'm pretty HP hped up on these Siri use cases so you know if you know what's good for you I would definitely recommend making at least one Siri template submitting that anyhow so that's uploaded duplicate that this weekend and build it out for my Android peeps I'm very sorry I don't have an Android but if it's possible in the Apple app there's probably like 10 different apps that can do something uh like shortcuts can on the Google side if you figure that out please share it with the community if you record a quick little loom video of how you do an Android I'm happy to share that in the next updates and show you some Lear what's next let's have a look I don't actually remember all this stuff guys you know I have notes all right okay and as a quick reminder this weekend the whole weekend I'm hacking along with marel we're at the AI hackathon Factory bill in and we're building drum roll a data scientist assistant what does that mean this is an AI agent that understands mathematics and statistics and complex product Telemetry and can crunch insights for you so you can see in my face as I talk about the scope of that it's Dawning on me what we publicly committed to try and build but it's the AI Sprint so we're going to get it done of course on the next Vlog update we'll have a nice little cut down of that but if you're curious to follow along as we're building at the hackathon I'll be posting more so on my Twitter and Linkedin throughout the day I'll probably also trying to get feedback from the community as we're building stuffs and there's another project that maybe didn't get as much fanfare but this was the research paper assistant that I worked on with OLG so it takes a academic topic finds academic papers on that topic and summarizes them that template is out live and you can download it as well it's currently on the project board I haven't uploaded it to any niio templates directory yet but if that's a use case you're curious and checking out that's available in the notion project and you know this is summarizing academic papers but really any kind of like big document summary that you have to do this is going to be a pretty good patent that shows how you would generally do that so go check it out even if you're doing sort of a business use case that has nothing to do with academic papers this pattern still shows you how you can basically search large swarts of text filter relevant texts and summarize them but yeah build build like just so you understand even though it's a very simple template the power of what it can do from the shortcuts app you can automatically get things like the location you can ask the user to take a photo front camera back camera they could take a video they can do an audio recording they could like you could remake Shazam right like what else you can also capture voice from the user you can show a modal and capture multiple form steps there's conditional logic in there depending on the response ask a follow-up question or for example you could maybe have a follow where like writes an email draft sends the email draft to the user they can quickly read it and then they have a button to save it as a draft or send it

directly hi Max how are you doing really well amazing job on everything you've done so far the velocity of everything and there was some surprising results like I ran it on a few sites and it was getting like 8 n firstperson emails across GitHub crunch Bas like all these different sites really amazing job man thank you to be honest if not AI it would not be possible and especially when it comes to the front end because I'm not like the proper coding developer but yeah great job on your site too especially when it comes to implementing caching and amazing materials that I hope people will see on product hunt so Oscar what I propose for our agenda cuz tomorrow is launch day on product hunt and we got a few things we still got to do which is probably pretty normal for a lot of product launches let me sh my screen and let's get into it here I've got launch toos how are you doing any blockers anything cuz I know you don't have a whole lot of time youve already done a lot to be honest no blockers on my side a few moments later so what's the potential if 2,000 people use it do you have any estimate of the cost it would be super cheap I assume so I don't expect to be like super like expensive thing because I've been running those scripts like for my own private like purposes for many workflows and I've executed it like hundreds of thousands of times to be honest th those bills were never high and hope they stay that way I think we can leave it on my Google Cloud platform and see what happens as mentioned I have hard limits even if consumption becomes crazy if we look at the to-do list it looks like I haven't gotten much done but I think it's my ADHD brain I have all of these things are half done so by end of day today this is all checked off and just expect for launch I'm coordinating that with Yan a CEO he's got a bit of a profile on product hunt so he's going to hunt this for us I've got your product hunt account so you and I will be made as the makers of it a maker it's very important the first message that gets posted as well I'm happy to handle that as well cuz I'm handling all the other messaging so basically at 9:00 a. m. when this launches tomorrow I'll make sure 9:01 that message is out there and everything and obviously from the NN team be doing whatever I can on promo a few minutes later and I'm just wondering which is more high priority which should I focus on because if I don't make both today I should choose something to start with great idea propose the CSS thing because that affects every single user that we use this experience the percentage of people wanting to clone the front end and the back end is going to be a smaller subset of that we I can add a note that hey the front end will be made public in a day or two check back here if you want it I assume the versel app isn't the thing that they're really excited to figure out how it works it's like a high scraping solution yes that day one so I think that's a perfect prioritization all right sounds good what if we look at the workflow and

you've explained at a high level maybe what's going to happen at launch we're going to get questions and we can next week a follow up or async answer some specific questions people had on the solution now maybe you share your screen and walk us through the floor sure we can do this that way make sure you hide all proprietary sensitive customer stuff before you yeah screen visible it's good now all right Oscar we're in the NN instance now where we're running emails could you walk us through this sure when it comes to the workflow that we have here we start with the web hook note set to the HTTP post method it basically sends us two data pieces the first one is the domain that are user wants to research and call back URL so the specific URL generated by the server where we can send back the data to actually display the emails to the user right here is the part that Max implemented before we even start researching we are asking superbase to uh check if such a query was made in the past right so in case like one user or many users quered the same website come a bit faster you know will be displayed a bit faster right of course when a result in a cach is being displayed to the user and sent to the front end when such result does not exist in a cache we have here the brave search it works similarly to the Google search we query the specific domain and next we have this fire craw mode which Returns the data in like markdown and HTML form so basically this handles for us the scraping of the page how did you find working with it would you recommend it to users when trying to do sort of a scraping or crawling use case so I would say it's very easy to use because this is basically like one end point when it comes to the scraping and crawling for this project we actually use only scraping visiting one website and returning it in HTML and markdown format but yeah API is super simple to implement forour use case specifically how did you find firecrow cuz I know you You' got some quite some scraping background so just compared to other scraping tools you've used over the years in that kind of so I would say it's like working very fine of course it has its own policies so when it's like encounters like LinkedIn page or YouTube page or any other social media it will instantly return you an error that is against the policy but for many other websites it works very fine I would say and especially for this further flow it's really great fantastic I think it's always curious cuzz you know this the marketing site of a page and then you use it on a real project and that's where you get the real insight on how useful a tool is for use case and from this one we have two paths the first one is information extractor and this path retrieves the email addresses using open AI model yeah then we have like this email extractor which Returns the list of email addresses on a page what is most important here in a prompt template we have instruction in this note to also include those non-standard email formats because sometimes those email formats are like for example in this format with ADD between those square brackets right so this will handle this thing for us and we have also the non AI puff which includes the HTML code of the page and we return here the data with a bit of JavaScript it's look not sophisticated right but in some cases it returns data which is also useful for the final output right simultaneously we are having here this AI agent most probably the most advanced thing in this workflow it we has two tools the first one is text tool which returns to the agent the whole text from the page right and the second one is your else tool which returns from the same page else in this case so in such a way this agent can go to this website to The Domain entered by the user and navigate through this searching for example contact page about us page and so on and retrieve the data from those what's important when it comes to those tools in this case here is the example text tool we are using here Puppeteer this one is not fire craw because otherwise we would hit rate limits very quickly and this is using Puppeteer which is posted on the Google Cloud platform right could you give us like a quick like two sentence description of what it is and how you used it so paper is a headless browser so basically you can instruct it what it should do on the website for example when it comes to the data retrieval in very easy words I would say so basically I have here two Puppeteer scripts and what it does actually in this case is visiting the website and is returning whole text from this page you may be wondering why we use for example papaer here and not like a simple HTTP call so the difference here is that when we do the HTTP call the JavaScript code is not rendered in many cases email addresses are hidden behind those JavaScript pieces when we use the proper browser Puppeteer in this case and we are visiting this website we can retrieve the data that is hidden and rendered by JavaScript in this case we increase the possibility of returning the data for our pooses a few inches later in such a case error is being like in the whole workflow and error workflow is being triggered and this one is sending the call back information back to the user so on any error that you have in your workflow the user will get a response uh yes there set the error work so okay guys this is very exciting as someone that like helps make an I haven't heard of someone doing that before that's amazing but what Oscar has done is because he's got this callback logic and he's sort of saving it in the workflow in the first step he's able to pass that along to an error workflow separate workflow based on any error in the flow and the user will get a response that's an amazing solution because otherwise if you use an error output Branch you have to predict what the error is to know to be able to send that back to the front end like in N end the eror workf for so fire and you as the developer would get that message but your end user wouldn't Oscar that's genius thanks so and then if it it's not an error case right so we've got the one on the top which is the Puppeteer version on the bottom of the fire C version basically they both go do an attempt to find some emails right in the merge node you kind of D duplicating that and making sure that you know we don't have duplicates of those and then we're basically doing some little transformation getting that data into one object and sending it back that that's yeah pretty much it and then the one last thing I'll add there is uh the last note in the workflow is the cach node where we add this new generated record to the cache now one thing I kept in mind at first I had that before we send the response back but then I had an error in the node which quickly made me realize I'm interacting with superbase it's a third party service it could go down it's far more important that our user gets a response than we get our cash set absolutely so I added that step on the end so empathizing with your users that's a bit of product mindedness can also help even when you're designing work absolutely so Oscar we got a few things to get done so thank you so much for spending some time and showing this workflow off to the uh Community if you're seeing this video and you have more questions Oscar is an ENT Community member please do ask your question CU we'll have an opportunity in future to go into more of a deep dive if there's a lot of interest to go into this workow I think there's a lot of moving Parts here we just went over the high level Oscar thank you so much for your time and here's to a successful launch of emails Ty oh to be honest I'm looking very forward for tomorrow so yeah fingers crossed and then I'm going to say this on camera cuz Oscar I said is going to get a flow grammar t-shirt I haven't made those yet after the AI Sprint I'm going to get two Email Spy t-shirts made how does that sound I will wear all right so let's get building thank you so much Oscar for all your hustle I really appreciate it yeah thank you to Max for inviting me for this like amazing Sprint it was like really cool there like past few weeks and yeah keep it going amazing cheers man bye 12 seconds later okay

let's talk model models and AI stuff so what's some things in the AI space that you're excited about oh so there there's the new family of quen models that just been released quen 2. 5 they have various model sizes so it's started from 0. 5 billion I think up to 70 72 billion and they have them for many use cases the most common use cases like Co coding math or general I this was just released yesterday so I didn't have much time to really play with them but they're generally very Stellar models although a bit censored also new mistal model just came out yesterday I think so the mistal small is like 22 billion model that's been available on the API and now the new version is also available as open source although the licensing there is not permissive for commercial usage but also seems to be very capable model according benchmarks although with m models sometimes struggle using them with the tool calling they don't seem to be performing that well for a multi-tool task where the model should be able to know that it should call a single tool more than once they often call tool once and then try to give you the response even though the task is not finished yeah new Cloud oppus Model should come out 3. 5 and I think that's going to be very interesting one commonality between them is they're all smaller models if we compare I don't know I think some months ago the headlines would be around things like oh you know like 405b how many billions of parameters that was kind of the big headlines I'm seeing now more it's not about its absolute performance on a benchmark it's the relative performance given its size right look this 7B model look how well it did it outperformed a30b model would you say that's a trend of getting smaller models that are a better at specific things how do you see that landscape oh for if technical means to train your own model for your use case that's definitely going to be cost effective compared to using some larger model that is more General focused but I still wouldn't use for example this 22 billion Mell model it's still not as good as some of the proprietary models like cloud and son so it definitely has limitations but if you train it for your own use case and you have enough data it could outperform this these larger models now speaking of fine-tuning something I saw recently with flux that fine-tuning it on 10 15 images of a specific person makes it a lot better at being able to render that person that's the smallest sample size I've ever seen from fine tuning I always assume you need hundreds and thousands and so the kinds of data where even if an organization has it you got to make sure it's clean and typed it's not really accessible for the average or the average individual but seeing this flux thing really challenged that for myself I can see you know even e-commerce brands for a specific type of shoe have models that understand these kind of products and stuff and generating them whatever but handle on what kind of if I'm thinking for my use case is fine tuning useful is there any guidelines on like if you don't have this level of data don't even really consider it handle on the kinds of volumes of input data to make useful fine tuning in most of these fine tuning jobs like you're not retraining the whole model so you just like for flux I think it's just a Laura adapter that you train on top of that model so just a few layers that's why you don't need that as much data because you're just tuning it and can you explain just real quick what Laura is I'm seeing this a lot I don't really get it nope me NE okay that's above our pay grade yeah but yeah you don't need that much data dat even the open AI for the fight uning they say it's if you have good 50 or 7 use cases that should be theoretically enough to see significant performance okay very interesting thanks a lot I really appreciate it I'm sure it's not going to be the last time I ask you some questions however naive they might be but I think a lot of us out here today have a lot of these kinds of introductory questions as we think about this space appreciate it my pleasure 6 and a half hours later when you're

building any of these workflows I've gotten some folks dming me with individual workflow help being honest it's probably the least efficient way for me to help you because I'm out here making these videos and FL gramming so if you get stuck on any template or anything that you're doing in an n Go to community. nn. it's the official Community forums we've got NN team in there developers and Technical people that can help unblock you we've also got community members around the globe who are super knowledgeable about NN you know they've been using it for years and that's nice thing about the NN Community is you see this when you meet NN community members in person well like I don't there's like that nice Earnest energy you know like we're just a bunch of helpful nerds that want to help other people plr if you get stuck in anything go ask a question there I recommend don't DM me two reasons my mental health and B if like it's B it's going to take a long time for me to respond relative to like all the awesome people on the community and see I might not know like I'm not an expert at every single thing at end and end but like we definitely have experts at all those things but the biggest reason is you know you ask me a question I answer it and I'm probably going to do that because I'm a nice guy but the thing is we can't Mark that as a public solution and others can't find that solution in the future our new AI co-pilot Ava which were triing out on EDI in cloud and then later will deploy to all users she gets smarter from every solution in the Forum by posting a question in the community forums you're helping Ava get smarter over time which helps everyone maybe not our competitors but you know so yeah I'm Max this is AI Sprint you're awesome one eternity later I I've reach my limit of people giving me weird stairs on the street cuz I'm walking down with a selfie stick camera so that's all for this update thanks everyone for your support wish us luck this weekend at the hackathon and we'll catch you next week for what's going to be some of the last days of the AI Sprint I might get a little sentimental I don't know if that's like emotional or just like exhausted exhaustion ex your so with all that mumble jumble of words being said I'm Max you're the Sprint Squad this is the AI Sprints and I'll catch you on the next update oh and how could I forget once again thanks everyone I wake up every day with all these excited folks all saying hey in my free time I set up Inn I tried it out there this one lovely message said hey man I've had nadn sitting on my computer for a couple years and your videoos got me to actually open it up and get going that's fantastic it made my day as you can imagine made my boss pretty happy too thank you guys making me look real good no but truly maybe you can see it in my face I'm not a professional actor right I'm just you know a flow gramar I you know I guess online with anonymity people can be such stakeholders so it's truly or inspiring when someone in their busy day especially some of you guys messaging me you're kind of big deals your heads are this and that and you still find the time to send me a lovely detailed message I live for that [ __ ] thanks guys w