Building Needle: Vibe automation, RAG workflows, and founder lessons

Hi everyone, my name is Mart. I'm from Assembly and I'm joined here by Yan from Neato. Um, today we'll just be talking about uh startups building with voice AI and Yen's experience as well um at Neato. Yan, thank you so much for joining me today. Can you mind do you mind introducing yourself for everyone here? — Yeah, also happy to be here and talking with you, man. I'm John uh from Needle um and uh we are building automations. So we want to make it possible that everybody can build automations. So we are basically a vibe automation platform such as you know vibe coding is quite of a thing to allow everybody to build like cool applications with code and lovable is maybe one of the leading uh products in this case or cursor. Um I think there's a massive new emerging market which is like automations. I think 2026 is the year of automations and agents. But to build these sometimes you may have to use a Python SDK or you use some Versel SDK or maybe you need to be relatively technical configuring JSON objects and the people who really want to build automations are or know how which automations to build are not necessarily the technical ones but the ones who actually are in the field and hence we want to give them a tool to be able to build automations for their use cases and so yeah that's why we created Needle um a VIP automation platform so with built-in rack and we also have an integration with assembly AI um which is actually also we do some hackathons so we started this 3 months ago and since then we're doing hackathons and we have around 50 participants normally in the hackathons and we also see that assembly a is a relatively popular uh integration in the hackathons which people like to use in specific case together with telegram because let's say you want to send a voice note on telegram and You want to trigger use this as a trigger and then you may want to use assembly A to convert this voice to text and give it to an AI agent because AI agent cannot work with audio has to work with text and then maybe do like a specific search across your Google drive files to find an answer for the question that person asked on telegram via voice and then may want to convert this back uh to voice uh um with assembly AI and then send back a message in telegram. So I see this for example as a classic use case with the seminary that's why I reached out and I thought maybe could demo this kind of use case together because I think it's also very relevant for like let's say customer support on WhatsApp, customer support on telegram or so and that's kind of like why I wanted to chat with you and so maybe we can see if there's some potential of collab or something like this. — Yeah, great. Thanks for the great intro. I um I wanted to say it kind of reminds me of like I guess the tools that we use internally kind of like um N8N and Zapia. Um but the difference kind of looks like uh there's a rag built in. So you're doing like an active search over uh a knowledge base. Can you maybe kind of tell me a bit more about that? Um I'm kind of used to working with Zapia or NN. — Yeah, sure. I mean there's two differentiation factors. So the first one is that as you said rack we started basically as a rack API and became a rack for builder. So our initial product was just rack API only because uh in 2025 so last year we saw many people want to build rack applications and it's relatively complex maybe because let's say you have to set up a vector database you have to think of your chunking strategy let's say your data lives in the email. Oh, you have to do a lot of work to in the end you don't care. You [ __ ] just want to share a customer support chatbot. You don't want to spend like uh two weeks on setting up and maintaining also the software uh for that. So that's why we were like starting with the rack API and we realized many people want to do repetitive tasks with the rack API which is for example answer the email you know so based on my knowledge base uh and instead of them manually going ahead and doing a semantic search to find the relevant ticket for example for this email uh we thought okay it seems to be like there's a huge demand for building for this automating with rack and so that's why we built then the automation platform And um yeah, so if you have ever set up a rack kind of workflow with other uh automation tools, maybe N8N uh I think N8N is a great tool. I'm not like thinking is a bad tool. It's super cool and also has a great community. Uh and um we're in a similar space, right? But if you have ever set for example rack and niten, I think you need like I don't know eight nodes or something. You need to you have some configuration issues with like the maybe use quadrat vector database or pg vector and then you have to think yourself of like actually all this technical stuff and if you actually just want to quickly and ship a high quality also solution that for example in the hackathon for example we had like with the you just want to have this telegram support and have this search kind of through your knowledge base which is maybe your zenes tickets which you have

just automatically indexed with needle and you can do record them or maybe your um Jira tickets or confluence documents um then you don't want to set all this up and you anyway would never be able to do in such a short time and so the overhead of setting all this up is too big that some people just let go and I mean you maybe follow also Reddit and so many people on Reddit I actually say oh I tried to set up this like rag pipeline in it and I have some problem can somebody help me out it's like massive amount of people asking this in different channels in the Nitn channel or the rack channel or the classic uh Reddit channels and I think it's one of the interesting sources also to know about what people struggle with and so yeah the needle is super easy you just have like one note which is needle note the needle agent which has access to then [snorts] a collection and a collection is a set of files that you connected let's say to different sources Google drive notion uh confluence and maybe slack depends on what you want to connect you decide And then you say okay agent you have access to this and we do have indexed the data. So we do rack on this. You can now use the rack tool do a deep search on this based on the input that you get and then give me back an answer and that's it. So it's just one note instead of like six seven eight notes and you don't have to configure anything. You just have to click click and then it's ready to go. So it's we often work together with like uh medium-sized teams um consultants um marketing agencies which are somewhat technical. They are liking to use technical solutions but they are not necessarily product uh software engineers. So in that case that is kind of like the suiting solution for them. Yeah voice is right now a big market. So you're doing you're in a great market assembly AI I mean we are not like necessarily voice we are just like kind of like uh helping to stick together solutions right like your solution is one of the parts of the total solution I mean Google drive for example would be still needed right but um I think it's an exciting market you are in and very cool also recently very hot uh miss 11 laps also I think they are they went pretty big uh right now they just announced another round I two days ago or something right 500 mil at 11 bit evaluation or something like this. So voice exciting market I think also voice agents voice customer support. So I think I will see more uh and more of these kind of workflows also built by users within needle. Um yeah in this — there's a fair amount of overlap like um we get lots of users building voice agents who need to do rag over like you know a series of customer service SOPs or u maybe they have a knowledge base. We internally have also kind of had that rag struggle of trying to like get our docs to be um searchable by voice agents. So we have a voice agent um and we wanted to answer technical questions about assembly AI. So we try and feed off all of our docs uh into like a database and you know sort of similar to the story that you were telling earlier about how you know setting that up can be difficult and you put a lot of effort in and you don't really get that great of a result but um you know it's cool to see that you know needle for example makes it easy for even folks in marketing or product managers to quickly spin up an automation and um you know uh give an agent access to their files. was within one node whereas in N8N it would be a bit more of a more complex setup with multiple nodes. — Yeah, I think we strive for simplicity, right? I mean it's a bit like maybe you have WhatsApp came out, right? And WhatsApp now is kind of the text messenger for everyone. But there was SMS before, there was like uh other services before. But it's just that the UIX and the UI, it was so easy to send a message to another person that this kind of like simplification of making a problem very simple and just understanding user behavior and creating a great UIX uh I think is like sometimes also very strong. Um surely the must also work well. The product must also have quite high quality. But I think understanding user behavior, user flow to then bring it like a solution which makes them feel good in a way. Um something very important, right? Because everybody these days does not have so much time and I feel attention span is also reducing these days. So some people just like I [ __ ] just want this to work like I don't care like make it work. I don't want to deep dive in a rabbit hole and learn about semantic search. I don't care. I just want this to work. So I see this sometimes as a technical person myself sometimes hard to understand because it's like oh it's also nice to understand deeply what's kind of going on and the architecture behind and stuff

but sometimes you just don't have time and you just need a solution and let's go you know then I also understand for example I don't know so well about your technology behind assembly I voice to text probably have some really cool uh models there and look cool training and stuff like that um but I just need a note in my workflow that converts my voice to text and so I mean I also would love to dive deeper I find this a curious topic but I also have so many things to do that I don't have time to dive deeper into this topic um so in this case I just use assembly as a note to do voice to text you know — and it's nice yeah — yeah it's like you know the there's um there's so many things to explore there's so many possibilities now with vibe coding you know all of us can use technologies that we previous previously didn't specialize in and that's also sort of a question I have for you is like um do you find more people are wanting to work or liking the user experience of working with a canvas interface to build automations compared to just using something like cloud code to like write a script or you know where do you see customers kind of finding their comfort zone because as a technical person I 99% of the time would rather just write a script with open code or something like And — yeah, I mean that's right. I mean uh it depends on the use case. For example, uh maybe you have a repetitive task that you want to do repetitively then and you want this to run for example every hour or every time a new email arrives for you to set up a web hook trigger then with like in your code might be a bit annoying to be honest. Uh so that's nice to work use a workflow then. Um or on another hand sometimes you like if you're not super technical you like also to see visually I see it a bit like Lego bricks and you kind of see the just build with Lego bricks something cool and you can see oh here my Lego brick broke my note is red so I need to fix here maybe I chat with the needle chat and needle chat is saying oh yeah I see your note is broken I need I will fix it. Uh people like this visual aspect I think because they feel like a bit more like I don't know what's going on. If you just maybe use code and write a script and you say okay Python execute my py file I you don't really and you have just vcoded that you have no idea to be honest what went on in this thing. Uh so you feel a bit like oh no idea how to fix this and then it's a bit less feels a bit less uh how do you say trustable maybe for some people like because they don't know at all what goes on and maybe then hard to fix also longterm and if you want to repetit repetitively run a specific task with a trigger note or sketch surely you can set up a chrome job also with code and stuff like this but I just feel like the human beings are somehow visual people we like visuals I think for example with Slack I think does an amazing job on visuals. I could also not have any visuals in Slack and just have a cold text messages. But some are really like this sound or this I don't know this da da or typing the other person seeing these things because human beings in the end visual people I think um Reddit does also an amazing job in this. I think it's a bit like this dopamine kick you know for example when you create the code and you just empty py file and then it doesn't feel like so it doesn't give you so much dopamine I think when it runs green I mean it gives you more dopamine when you see the small boxes turning green and it turns green all over time and then oh every turn green oh now it works that's for even personally me myself I'm a coder right but sometimes I just like to use a workflow because I don't know it feels more it feels Somehow for some tasks it feels nicer to use a workflow. I don't know. Yeah. — Yeah. I mean I Yeah. I I can see how that would be um it would kind of it's easier to feel a reward. feedback from that kind of like — thoughtful UX and thoughtful UI. You know, you can tell when where exactly what is broken. Whereas in code, it just sort of feels like a big mess that you don't even know how to navigate. I I I definitely see that. I guess that's honestly a great segue to um the next question which was can you show us around Neo and like show us some of that UI and that UX? I'm curious now after I guess hearing your thoughts about thoughtful design. — Yeah, sure. Uh cool. Uh so I'm here with a needle uh and so we are in the needle kind of like uh space and we have different kind of departments and so for different departments we also try to create different kind of projects and for example I could go here let's say to uh the marketing project and then I have kind of like uh different workflows and

I could create a new workflow and let's say I I could say I either use a template whichever template I want to use or I start from scratch. Um, and I can create like this workflow here. And now I'm here in this empty thing. And now I can go on the right side. And I don't know yet what I want to build, but I kind of know what problem I have. I want a workflow that runs every day at uh 900 a. m. reads my last 24 hours uh Gmail emails and uh summarizes them with a agent then sends to # generate slack. So you may do like this in vibe coding too, right? You would maybe tell this to cool code or something to like this. So what I wanted to say is like vibe automation. So in this case I describe what I want to get built and the agent goes ahead here you see and then is actually building this workflow for you on the left side. So it's like vape coding but we call it vape automation and so you don't have to you say yo looks good uh sounds good to me I have done this and then you see schedule trigger every day at 9 as I kind of like uh described it. Then we have here a prompt uh a Google mail note. We have here an agent node and it's kind of like you prompt just everything. You don't need to maybe I would like to use a different model that could be chooses for mini which is maybe not the model we want to use. So yeah it just creates this workflow kind of for you and you don't have to manually go ahead. You can also go manually go ahead and now I could here for example go to my assembly for example uh transcribe audio note and let's say for example I got an email which we where we say we will say find email I don't know that has audio inside then I could pass this to assembly and transcribe the audio and maybe then I would go to the AI agent and say hey you transcribe the audio and now I can give my agent to it. I could either say for example you can search also the internet and maybe I would also search uh my collection. So my collection I need to decide which one I want to say. Uh for example I maybe want to search the Epstein files. We have like an Epstein file collection. Uh and you could say for example yo you can search internet and search collection to answer this whichever comes from this. Now I would need to connect to assembly in Google mail which I can currently done but yeah basically you are done. uh have created kind of a workflow for this. But what I could also do if I would have created this connectors, I could give this all the as tools which I don't have yet the assembly connector inside here. But I could also just give this all as tools to the AI agent and AI agent would itself you describe something like uh first transcribe audio to text with assembly. I would need to select assembly as a tool. So it's kind of like you know cloudbot but with a specific set of tools it just has access to. You decide which tools it was access to. I would need to create a connector and then it would have access to and then I wouldn't need to use the assembly node but I would kind of put the assembly tool to the agent um and it would understand first I will always use the assembly tool. I will transcribe afterwards I will to cause the search web tool or search collection tool and afterwards I will generate the answer for that you know so it's kind of like you could either put everything into one agent and let the agent call tools or you can just also visually see the boxes here and maybe I can go to one workflow which I already have set up um LinkedIn from Zapia find LinkedIn list from Zapia this is like a deep research mode. Um, track define competitors promptation trending content generation. H I don't know which one I should run right now. I have like many different ones. I don't know which one is interesting to be honest. But yeah, I think I've shown you the UI. You can like invite members to your project, work with them together on like different workflows. And one thing is cool is this collections part. Um I think the thing is that you need to create collections and so you can create new collection assembly AI for example I hit on save and to this collection you can uh you can then go ahead and you can kind of put create connectors and this is what I come to the rack part. So the collections are you can create connectors here. So you can make this collection available to have access to HubSpot or maybe Google Drive, Jira, Confluences and notion as I

mentioned Slack. So you make this data rack available. But you could also say for example all pages from a website and I could for example take your website assembly AI you and I could say I want to put 800 pages you have on your website. I want to create an assembly AI connector. Um and then I could give this as context as I have shown you before in my workflow to my uh agent and say search collection assembly AI. I could say create this connector and once I create this connector it will take some time. This is 800 files and then afterwards here we have a chat. You see you could ask this chat questions or you use this chat as I have shown within the AI agent in a workflow. So you could have for example let's say you have a customer that ask a question about assembly AI. you have an FAQ page somewhere on your website likely then you would just uh use this collection and you would say yo um just uh create like uh if this person questions uh Gmail look at the assembly I collection do a rack search on all this data which comes here inside now we see there's a lot of like data coming in from your website and generate an answer based on this content that you have from this collection so we have like an automated customer support bot which is bas on your assembly data having rack already built in uh super fast. Um so I could go back to my workflow. I created this assembly AI workflow. I can go to my agent. I say assembly AI now. Uh I can choose this collection. You see I say don't search internet. Just use this one and then suck. I have your kind of customer support with built-in rack on your landing page data ready. You could also connect like Zendes tickets or whatever else you have to answer these customer support tickets for example. and super easy to be honest. — All right, so that was a great demo. Thank you so much for showing us around. I um I liked how easy it was to use. And um I did also get that kick that you were talking about where you can see, you know, some of these bricks, they're yellow. Let's go in now and fix them. I was uh that was great. Thank you so much for sharing that. — Yeah, thank you for letting me share also. Yeah. — Um, so now I guess I kind of want to talk to you about being a founder and building something, building a company. Um, maybe you can start by telling me like how long you've been at Needle and um, I guess the origin story where you started. — Oh, um, oh, origin story. Origin is a hard question. No, because like where do you start? When you're born. Uh, no, I mean, just kidding. Um yeah, I always wanted to start a startup and I just thought life is a bit boring. I think always I find life a bit like not so exciting. Uh I studied for example two bachelors at the same time because I thought one is not challenging enough. Maybe that's a bit weird and I wouldn't recommend. And then I worked for some time in San Francisco in a YC startup and I thought this is really cool and great spirit and I love this hustle and drive. uh live for some time in Switzerland and work there as a software engineer. I felt like life is a bit too predictable. Like it's like okay, you have a nice job, you have a good salary and life is good, but you know, I feel like life is meant to be more than this. And so I wanted to create something great. I want to create something which excites me and also excites other people where you can which is like I don't know like passion I guess. And then I created Needle like um together with my co-founder Uno uh around one year a bit more than a year ago. And why did we start it? As I said, we started as a rack API because I had many friends who are software engineers, consultants and so on. And they told me yo they may work for Accenture or Deoid or something like this uh or some smaller consultancies and they said everybody nowadays wants to set up rag pipelines inside our company. So we are trying to use different tools and we try to understand what can we do and I thought hm everybody wants to set up rack and it seems to be like a bit cumbersome why is there not like a rack API for that like a stand out of the box solution such as maybe in the beginning you would set up websites yourself like manually or databases but then there are managed databases there are managed kind of applications that let you manage websites and so on and so I saw the same is happening with rack so rack infrastructure tool um which is the case. So, rag MPI is exciting and in demand but also became a very crowded market as you see like many players joined the rag API market. We were very early I think back then. Um and then we as I said again we realized like people want to automate stuff. They don't want to just do the search like let's say cool I index your assembly AI website and I can do a search on this and can ask questions about your pricing or whatever. But you now need to put this into like a real use case. Normally rack is not just used standalone like yo I use rack and so what I use it in a

business use case which is like customer support for example or um maybe I have a team update or I want to read all my Slack messages and every day give like a summary of what happened on Slack today within the team you know so you need to bring this into a business context and that's where automations came to play very well along I think with rack and that's kind of like I think the take and we raised the 2. 2 2 million round. We are a team of like six now. Um we are seeing a great uh drive and uh people very like excited. Um and yeah, we are trying to make this like 2026 year count very much and um yeah that's kind of I think it but I mean story is I always was excited to start a startup and I feel just like bit bored if I wouldn't do a startup because I feel like I think life is something life is not shouldn't be that predictable. I think I like that it's unpredictable startup times, you know, every day is kind of exciting and it's not like you work. Yeah. — Nice. Yeah. Thank you for sharing. I think um I think that resonates with a lot of people like uh what you said about being passionate about something and working on something that you really believe in. I uh that's that's really commendable and you know I think uh it's really great that you are in a position where you can live that out every day. Um, yeah, — you know, you're building something and you're working with customers, you're giving them something that they want. — You know, you pay all this attention to the UI and the UX. Um, that makes a real difference and I'm sure it gives you fulfillment and um, at the end of the day, I guess you you're happy and um, that's great. That's that's what life is for. — It's the hard days, you know, not every day is fun. — That's right. Yes. — Yeah. Yeah. How about you share um some difficult times in uh in being a founder? — I think last year in July or August. Um so we launched this Red Flows last year in October actually and we made you know sometimes it's like going to gym you need to [ __ ] lift the heavier weight. Uh and so we had to lift very heavy weights because you know the problem the tool we built look at the product is a very is also a complex product. It's not like we built a marketplace or Shopify store or as you know is a you know with the chat with the rack and so on is a very deep tech complex product and to make then on top of that like a good user journey is also really hard. Uh, so we really had to go very much in deep mode like hustle mode like crazy. I mean I'm still very much in hustle mode. I guess I stayed in office until 9:00. I'm working on the weekends and stuff but uh I think back then in July I was uh I think I was close to burnout to be honest. I could feel it because I was actually I had many nightmares. I could like not sleep so well. I was very stressed. Oh, I think that was actually a hard time. Kind of came out of that good friends spending more time doing stuff with friends after work and trying to deconnect a bit, disconnect a bit, you know, sometimes otherwise if you, let's say, take your laptop into bed and you work until 1:00 in the night with your laptop in bed, then your sleep is going to be very bad and next day you're very tired and then again it's a safe enforcing effect for the next night. Um but I think sometimes you also need to go through these phases. Startup is not just like yo you ch 9 to5 and then everything is going to be great. You have to do this extra push because competition doesn't sleep you know. Um other people also push. So I also believe in to work smart and not hard but you also have to work smart and hard at the same time. I think there is not just like only work smart and not hard. So that's kind of was a hard time. I think I was very down. — I think the good thing about hard times is that you come out of them stronger and um you know that's why that's part of like the journey of being a founder is like I'm not a founder. I've you know I joined assembly fairly early on. Um, and so I've seen people who have been here for really long and um, you can tell that they've been really kind of shaped by the struggle and um, and that's like that's all part of the journey. And um, when you look back on it, I guess you have happy memories of uh, of the tough times because like you said, you came out of it with friends and um, with lessons about yourself. Um, another thing that you mentioned was that you solve like for your own pain like you had friends who were talking about

building rag internally and um you figure that you might try and see how you could help. Maybe you could kind of tell me more about that discovery process of finding a problem to solve and um how you landed on rag as an API. — Yeah, I mean ra as an API was the very first tool. Uh so now we are refloa which is again listening to users right but I mean it's just like I was software engineer myself in a company where we wanted to roll out something where we need rack API and then if your friends tell you the same I'm like yo everybody working now on this seems to be like a market uh then I followed a lot Reddit I think Reddit by the way great tool for market research because people are honest on Reddit they don't like lie they are just like uh telling their problems they are saying what they were built and I saw there's so many people who had this problem and then uh it was just like I would say that's how we validated it. Then we try to ask some people hey would you use a rack API or we say if we have a rack API you have to spend less time to set the rack would you use it and people were yeah I would use it and then that's kind of how we validated first people trying the product using the product there was demand for it and then we just said let's [ __ ] go you know I think sometimes you also say need to say let's [ __ ] go because you cannot always just say maybe let's stay a bit more inside and think a bit more at a certain point I think you need to say I go out there I do stuff I push and I try to make it uh because if you never say let's go then you can also never make it but uh yeah it's good to validate beforehand and work on something that you like but I think definitely very validated uh problem space um but I think also you don't know so much yet in the beginning what exactly solution you build you know kind of I like this problem space it fits to my skill set co-founder skill set. It fits to my what I like to work on also and it fits to the problem that I want to solve and the clients serve because in the very beginning we were working together on something very different also which is a point of uh service POS system something like sum up you scan a QR code on a table and you pay and I talked with like 200 restaurant owners and I hate to talk I mean I don't want to they are great also but I cannot so well talk with restaurant owners because they way of thinking is sometimes a bit like old school and then as a young technical person it's somehow very hard to resonate with them and if these people are your clients and you have to talk with them every day I would have burned out after half a year and I would never come back to this you know so you need to love also your user you need to like to build something for your user and I think in our case user is relatively similar to me or like people I like to work with for or you I think you could be user I think you're a nice guy also. So I would like to work with you for example, but I would not necessarily like to work with the 60-year-old restaurant owner who is very much against tech and uh a bit like conservative. I could not work so well with him every day and see him every day but still be like smiling at him say, "Oh yeah, I surely help you because he's his user, you know, a client and client is king. " — So then very fast. That's funny. It's kind of like you want to work with your friends and have a customer that you like you genuinely like and genuinely want to get along to talk to every day. — Yeah. Exactly. Because otherwise you see them every day. You talk to them every day. They tell you their problems every day. You have a buck and they come to you, — you know. So you need to like them like your user. So it's like the problem you solve. And I didn't know that we would build automations back then, right? But now we do. So, you always need to be adjustable to the market. You need to be fast. smart and hear what people say. — Um, if you had like one piece of advice for it's it's a great segue, but if you had one piece of advice for founders or product managers kind of watching along, what would you give? — To be honest, I think find a great co-founder. I think that's very important. — Oh, wow. Interesting. What's your relationship with your co-founder like? Oh uh it's very good I think I mean we [snorts] also have days where we are not where we disagree because I think as oftenly founders have uh strong opinions and also some ego. I do have some strong opinions and they do have some ego but you also need to have because you need to move fast and push and you need to be believe in what you think is right. It's also okay to question yourself but you also need to be very confident and sometimes if two founders are very confident they can also clash because it's hard then to find a middle ground. I think uh if I first also thought maybe to found alone but I think it's very good to have a co-founder to because

some days you're low the other person is low but uh you are high so you kind of push each other you know you push each other up because let's say you stay in the office on a weekend on a Saturday until 8 and it's dark outside it's hard to do alone but if you're sitting together maybe an office maybe you order lunch or dinner together then kind of enjoy it more you know you also need to enjoy the ride And finding a good co-founder is very hard. Uh I think you need to put yourself into really tense and hard situations and to understand can I see this guy on a Saturday night while I'm not in a good mood and we kind of have some problem in production. And if you can say yes to this question then this is a good co-founder for you if you have complimentary skills also and they fit to the product. Yeah. — Nice. Thanks for that piece of advice. Honestly, thanks for coming on. I really enjoyed our chat. I enjoyed um listening to your story. I'll be sure to share like your links in the description. And yeah, everyone, that was Yan from Needle. Thanks for listening. — Thank you very much, everyone. And thanks, Matt.