Studio Update #03: Barcelona, Gmail Draft AI Assistant, VAPI Restaurant Voice Receptionist AI Agent

another week another Studio update hey everyone it's your favorite flow Gramma Max and this week I'm walking the streets of Barcelona and literally right there a little bit of gy I'm in Barcelona for team offsite officially the goal this week is to team build and strategize a global takeover but between you and me my personal mission is to ingratiate myself with my boss Luis on let them know all the good work that we're doing here and you know hopefully get the budgets needed to take a bunch of flow grammas to Greece put in it in space know they other very important things no one knew they needed but you definitely need this in your life trust me there Lis also build this as a sunny getaway in Barcelona but it's see it's 13° and 50% chance of rain brothers so we

giv as always let's kick things off with some studio Updates this week I was hard at work shipping the part two of the building AI agents tutorial series that's live now on YouTube should definitely go check that out in that I cover setting up AI tools for your AI agents these are the new tools available that make it super easy to let your AI agents interact on your behalf with apps and services and get all the kind of contexts they need to do really powerful stuff and also create data in your systems in that tutorial I show you how to build a Gmail draft assistant so it takes new emails that come in your inbox and writes a draft and then one key focus of that video is how to set up guard rails right what that means in Practical terms is how to make sure that your AI agent isn't deleting emails when it doesn't have any reason to how to make sure that it you're giving it the permissions to do exactly what it needs to get the Tas done and nothing more you're just exposing yourself to potentially more risk because I think there's a lot of people in the space are saying AI technolog is amazing and it truly is amazing but it's not perfect no system is and any engineer knows that so as part of that series I'm not sure what I should do next please do leave a comment I could go deeper into prompting we could focus more on structured output perhaps what one sec my girlfriend's calling me hello so I'm on camera right now and as I collect some ideas and inspiration for what you'd like to see in the building AI agent tutorial series next I may put that project on pause because I'm doing all these use case interviews with people building all these really cool AI agents so I kind of want to get my tee stuck in and build some really cool agentic use case and publish that as a template and then maybe get back to the series and so Angel has been hard at work on releasing the first episode of his nadn at scale webinar so definitely go check that out that's out now on our YouTube links to follow but Angel's also been a bit under the weather last week so hope you get better soon mate and he said he's going to pick up the to attack AI assistant

once he's back and ready next up I had the pleasure of interviewing Alexandra spalato an AI automation engineer based in Madrid she walked me through a fascinating solution where she's created a voice AI receptionist that can handle restaurant bookings now you'll see in a second but this thing works we were on a call together and she booked a restaurant appointment with a digital voice she built it with vaki and Ann so check this out hey Alexandra how's it going hi fine really great to see you and thank you so much for coming on the show I appreciate it thanks to you for inviting me absolutely before we jump into your use case would you mind introducing yourself for everyone I'm Alexandra palato I'm originally a web developer I've in Madrid Spain but I'm French and I love building stuff I have a YouTube channel now and the important for it's to build things whether it's with code or no code and so that's why I find N1 because I can use code and no code and make my things more efficient I've been a developer relation like you for two years but now I work a lot in AI space and AI automation space very cool well building it's definitely speaking to my heart I'm a builder at heart as well a tinkerer I have it on good authority that you going to work for to show us today could you explain the use case at a high level and what inspired you to create it the use case it's an AI vocal receptionist for restaurants so I live in Madrid is a city I think that has the most concentration of bar and restaurants in the world I'm not sure but it's quite amazing and I love going out Etc so when you call you cannot have people on the phone and you cannot reserve your restaurant Etc so I think it's a very useful use case and that can be declined also to other Industries so I build the AI agent with vapy but this agent need tools to check the availability of a table and then to reserve the table and also to I do it with her table with real table management because all the tutorials that I saw they were using the Google Calendar and that's not realistic for a restaurant because you don't have one table one people so I bu that and so we have three workflows in na10 that connect that trigger by vapy by the conversation that sounds fantastic so it sounds like we you've got vapy it's handling sort of the it's like a digital receptionist right or something like that it's passing the human language into things like the date and whatnot and then and ATN is handling the interacting with air table your source of Truth understanding if things are available if not and then interacting with vapy to capture the correct information it needs to then like create a successful booking is that right exactly that sounds wildly powerful would you mind sharing your screen and showing us the use case yes it's not only the screen I'm going to call the restaurant so we can reserve shall we book a reservation yes it's restaurant Fantastic San Francisco a bit far away for both of us Alexandre where are we right now on your now here we are in vapi so do you prefer that I do the demo first or I explain let's get the fun part let's book a restaurant reservation with AI and then if you could walk us through how that happened but I think everyone show how it works in a table first here for example we have tables the one has as two seats Etc so we have a table management small one here but it works and then we have a table with reservations these translated here to the table availability so tomorrow we are on the 10 so here I have made this unavailable and I'm going to reserve here to see how the assistant reacts that he can handle the thing let's do it and we will see here the reservation appear so I'm going to call this restaurant let's get us a table joural this is Sophia from Little Italy how can I assist you today hello I would like to reserve a table please and be happy to help with can I have please Alexandra nice to meet you alexand how many people and at what time and day would you like to reserve a table six people for tomorrow at 7 p. m. moment it looks like we don't have availability at 7 p. m. tomorrow for six people Alexandra however we do have tables available at 8:30 pm and 9 do either of these times work for you yes 8:30 p. m. it's perfect may I have your email to finalize the booking any special requests or dietary restrictions we should note Alex gmail. com and we have a friend which is Glon allergic alexand Alex gmail. com yes everything noted let me finalize your reservation one moment please one momento your reservation is confirmed for six people at 8:30 p. m. tomorrow Alexandra we've noted the gluten allergy as well if you need anything else feel free to reach out a wonderful day and we look forward to seeing you at Little Italy thank you bye and that's it and now we are going this is going to become un available and it'sing every minute but it's going to appear on the screen first off Alexandra I'm mind blown so we have with an AI agentic solution booked a reservation it handled paring I mean the gluten allergy thing even you said it a little bit differently from how they transcribed it was still the correct concept would you mind showing us what happened under the hood to make this amazing solution happen yes of course okay so this is the a table by the way Cloe helped me to create that because it was quite a headache to think all these table management the number of guest and then imagine when you book for example at 7 p. m. you need to make the next four slots of 30 minutes unavailable because the dinner is about two hours if we begin in vapi okay so in vapi we build assistant we have a first message and a system prompt when guest ask for reservation use a check table availability to verify availability okay and then if available proceed to collect email and special request then use Reserve table so this check table ability and Reserve table are tools then we give it instruction to how to handle the call by the way I have a document where I have all the automation I do with everything to reproduce them so I can share a link at the end of the video too um then here we choose uh the model I try with the mini but this morning I had an error so I put the 4 Z but normally should work with the mini we also have a knowledge base so all the information about the restaurant the menu Etc if people ask it's taken from the knowledge base I put the temperat to 0. 7 to keep it creative but not too crazy and then these tools then okay the transcriber is deep ground the voice I chose this one Jessica she sounds really well from 11 Labs here in the functions this is where we have the tools so the check table Ava and Reserve table and for that I go here in tools and here in integration This Server URL is the one from na10 because we are going to pass this to make the post request to na10 to treat all that then we give a name to the tool and also the description so that helps the agent to know when he needs to use this tool and in the here in the function we give some properties so we want the date the time the guest so this is a check table availability and the reserve table here so this is another URL from the other workflow okay from the other web hook and here yes we take the date the name the time the email the notes and the guest and the phone number okay so this is the check table availability workflow if we go to executions we can see this is this one and we can see that he has run to here and if we make it bigger we can see say respond alternative times because the requested time was not available okay so here in the hook so it's important yes to using the respon to web hook node because we are going to send back the information okay and here we have all the information so it sounds like you're consuming this tool basically as if it were like an API endpoint is that right yes it's a web hook so we have a post here and we receive the data from Ving and then we make all the workflow so first we have the tool we see that the function it's check table availability so we're good and we go here so here I passed the date with open AI to transform the natural language like tomorrow next week or Tuesday you know in real dates so I have a prompt here okay for that as I say everything is I will share everything at the end then we have lot of codes so I'm a developer but I didn't write any code there clo write everything from me so as in a table I put the 24-hour format and in the US it's the 800m ETC then I transform it here with code this one the last one here availabity on request at time was hard and there is lot of code and I would not love to write it so you have really to think what you have an input and what is the output and try to prom Cloe to understand what you want and this is this one for example was really hard because we have to check the four next slots to see availability for example you reserve at 7:00 p. m. but it has to be available until 900 p. m. so four slot of 30 minutes it was failing because if it's the end of the day there is not four slots there is only two because it closed at nine so then I realized oh I have to ask it the non unavailable slots instead of being available and then he understood because I tried model the Z1 preview Etc and he was doing it wrong so you have really uh you know adj just valid time to include times where there are not explicit unavailability in the next four slots so it's really thinking what you want as out putut on all the parameters and asking clae how to do it and clae knows very well in A1 so he knows it very well here so we pass the date we pass the time here we get the number of guests here we write code again to get the suitable tables for a given number of guests because if we have two guest we can put them in a table of two of three Etc so it check that again with code again this one is simple no but all by clo uh then when we have that we filter where we get all the available slots in air tables and we filter them by date availability and suable tables and here too this is a filter for air table I didn't know at all the air tables filters but so here okay so Claude came up with the air table filter Syntax for you there is that right uh yes I saw that and I asked CL again okay I have to do that and so CL wrote it but I didn't write the code so it's doable for anybody that can to take the time to really think about it here we have the tables that are available at this time uh for this number of uh as of this day and this number of guest but then we are going to check the time after that and here we check if the requested time is available on the slots and propose alternative times in the same period dinner or lunch so we get the requested time the available slots when the lunch hand when the dinner hand if it's lunch period or dinner the result is not available for the requested time but we have these alternative T we have if here so if the time is available we go here and so this is the structure we need to send back to V result we get the tool call ID this come from the hoop for the first call and we return okay if it's available through a phase here it's fails so here we have been to this route so here result is the alternative times which are 8:30 p. m. and 900 p. m. and so we send that back to VY and she answer oh it's not available but we have a table at 9:00 p. m. that's fantastic and my first takeaway here is how like right tool for the job and what I mean by that is you know vapy is handling human language via voice very fuzzy lots of ambiguity and it's able to turn that into specific data but then in your workflow you're using a little bit of AI to conform that even more to have very predictable inputs but then you're using non AI steps to do the predictable business logic some of this the availability requested time you know that is business logic it doesn't make sense to run that through a large language model take 10x the power and time to calculate that versus having a code node that you can inspect and say great this will deterministically run that way every single time and this is I think a great example how you're like picking the right tool for the job and not just using AI on every single problem and I think I see a lot of people in this space just applying AI bet trarily but you show because it correctly as much as we can stay deterministic and have you know the right Reas for sure it's better as much as we can drive the data to Ai and just need AI for his power and when we really need it it's also probably going to be a lot cheaper if you're running this in production right so it's now determined whether the slot is available or not what happens next she come back and said okay this slot is available do you want to take it yes if there is no slot available we have not seen this case she will propose another day do you want to another days so if the customer say oh yes I want to go for Thursday then she will do the cycle again and check everything but if the customer wants to reserve this slot then she will call well it will call the reserve table function so here again we have the web hook so if we go to the execution it runs here so here we can see that the function that is called is Reser table care for tomorrow uh my name you know take all the information get guest with WR allergy guest six Etc so let's see then how it works so here I edit the fields to get you know all these data to extract all these data from the call and then this is the same the switchable tables the par date the formatted date it's almost the same things par time and here in air table I record the reservation so I use Creator update the reservation table and I map all that to make my reservation and after that we have another workflow that doesn't interact with vapy but that updates the table availability table in the restaurant Bas so this one this is the air table trigger that is trigger when there is changes in the res reservation table so we have seen that we have made a reservation we get the data from the reservation table here from the changes this is the reservation we have done then we edit the start time table five the date and the reservation ID then here we create the additional records because we reserve at 8:30 p. m. but we need to cover two hours but at the end of the day don't take reservation after 900 p. m. we have the reservation the start time the table ID the date we generate four time slots Alexandra first off thanks so much for walking us through the details on this I think it's wild you know there's some complexity here but there's two tools a utility flow an air table then vaping but how it's relatively simple for the immense utility that's being offered by this so I'm curious what would you like to add next to this use case if I do it for a real restaurant for example there is many things that we can do we can use perhaps database like super base or something like that more robust to a bigger restaurant Etc build front end graphic user interface I'm a frontend developer so I can make something really nice we can even show where the table are placed and all that we could take the payment methods to set up a no show fee if people don't come it really sounds like from my side it's this is like a weekend couple hack days away from being like a micro SAS that's launched that restaurant to clicking the buy button for and setting it up and using it for handling dozens of hours potentially week of work so Alexandra again super impressive stuff thanks so much for showing off the use case and I'm curious why did you choose n8n when you were building this house because I'm a developer and everybody was with make Etc I try make and I cannot have code there and so it's very limitting so I discover n10 and it check the boxes because I can have code there is the AI agents that are really cool too and I'm going to make videos about that and to Dive more into that also is self-hosted so it can be almost free if youself hosted it so that's really cool so yes it's flexible it's open source the vibe is very modern I love the design I told you thank you kindly that's music to our is as you can imagine Alexandra I just want to say thank you so much for your time for your inspiration on this use case I've seen a lot of folks doing the whole vapi and N at end thing I think I got to go play around with this now so once your templates out I'm very excited to check that out and you mentioned that you're going to make this public where can people follow you and see this use case and future builds that you do so on YouTube I'm posting every week so I'm preparing a new video which is about getting ideas from YouTube comments to make other videos and things like that I'm on LinkedIn too I have well everything is YouTube and Linkedin I have to build my agency websites and many things I'm building a SAS I'm doing 100 things so busy Builder love building that sounds very familiar yeah that's the thing whatever is the tool I love building things well wherever you're watching this video go ahead and follow Alexandre following her awesome Journey because I you're building some really inspiring use cases I think also for our more technical folks I think the nice thing about NN is you know some people could have used less code nodes some people more I wouldn't say there's a right or wrong way with that right you can go a more low code approach if you're more comfortable that you can go more no code approach you come from a coding background you're more comfortable working with syntax I think there's nothing wrong with syntax it can be a very efficient way to execute and express yourself so thank you so much again for showing it and uh I can't wait to have you back on the show again so looking forward to having you back soony thank you

bye thanks so much Alexandra that was wild wasn't it definitely go check out her Channel show us some love she's also publishing the thing she builds publicly so you can download it duplicate it freak it tweak it and learn from it since I'm in Barcelona that's all for this week but next week should be a super interesting one cuz we got lots of updates from the NN team lots of cool features there's some exciting things happening in the code node and a lot of functionality actually around nen running more efficiently at scale underload yeah so I have some updates on that some more interviews and a sneak peek at what's happening next in the projects I'll be building I'm looking for inspiration I'm planning q1 I want to do a lot of different builds and interviews and that sort of thing what do you want to see more of drop a comment let me know and happy FL gramming everyone we'll catch you next week Che