AIMinds #013 | Peter Dun, Founder & CEO at Feathery
About this episode
Peter Dun is CEO and Founder of Feathery, the most powerful form and workflow builder on the market. Feathery also offers natively-integrated document intelligence and voice AI products powered by Deepgram. Before Feathery, he studied computer science at Stanford and was a growth lead at Robinhood working on user onboarding.
Listen to the episode on Spotify, Apple Podcast, Podcast addicts, Castbox. You can also watch this episode on YouTube.
In this episode of AIMinds Podcast, Peter Dun explains how Feathery is upgrading the form and data collection industry by harnessing the power of AI. He gives insights into the integration of AI in form and workflow solutions for optimized user experiences. Learn about the 'human form' digitization process and the future plans for Feathery in the fintech and insurance industry.
Here are some takeaways from this episode:
Form Builders for Enterprises: Understanding the need for complex form builders that deliver high level requirements and logic, Feathery caters to businesses like Robinhood, which calls for elevated data collection and user journey requirements.
Integrating AI into Form Building: By integrating Document Intelligence and Voice Intelligence into their platform, Feathery enhances traditional form experiences, enabling enterprises to capture unstructured data seamlessly.
Pushing the Limits of Document Intelligence: With a dedication to achieving up to 99% accuracy in document intelligence, Feathery is optimizing user experiences while maintaining the speed and accuracy of data extraction and processing.
Fun Fact: Feathery was inspired by the limitations Peter Dun encountered with existing form builders during his time at Robin Hood. They decided to build a form builder tailored for companies like Robin Hood and various enterprises with high-level needs for styling and logic in their digital forms.
You can also check out Part 2: A live product demo of Feathery on Youtube!
Show Notes:
00:00 Robinhood onboarding process for investors.
03:52 Robinhood onboarding flow essential for company performance.
08:24 Challenges in building Robin Hood's onboarding flow.
11:33 Complex onboarding process for financial service companies.
16:08 Feathery: form and workflow solution for businesses.
17:13 Automating data validation crucial for document intelligence.
20:42 Voice AI automates data entry and validation.
26:10 Document processing involves multiple important components.
26:50 Wrap up.
More Quotes from Peter:
Transcript:
Demetrios:
Welcome back to the AI Minds podcast. This is a podcast where we explore the companies of tomorrow built AI. First, I am your host, Demetrios, and this episode is brought to you by Deepgram, the number one text to speech and speech to text API on the Internet, trusted by the world's top conversational AI leaders, startups, and enterprises like Spotify, Twilio, NASA, and Citibank. Today we are joined by none other than Peter, the CEO and co founder of Feathery. How you doing, man?
Peter Dun:
Doing great, man. Thanks for having me.
Demetrios:
So I want to get into your story a bit. You were. You moved to the Bay area, originally from Indiana, started going to Stanford, but then a little thing called Robin Hood came along, and you were like, I'm going to start working there. You were working at Robin Hood as a growth engineer. And did you finish Stanford, or did you do the cool thing and drop out?
Peter Dun:
Yeah, so I was actually a Stanford for all four years. You know, it was a great experience. Want to be there with my classmates. I was actually kind of doing, like, a dual, like, uh, working at Robin Hood and then kind of, like, living on campus and stuff at the same time.
Demetrios:
Okay, nice. And were you at Robin Hood when there was the shakeup and all of the fun stuff that was happening?
Peter Dun:
Oh, yeah. I feel like it was kind of Robin Hood for, like, the. The exact period of, like, hyper growth where, like, all the crap was happening. Like, I think, like, first week I was there, we rolled out our high yield, like, cash management products. Two days later, the president of the SIPC came out, announced that we weren't insured. We got to roll back everything. When I first started, and then a month before I left was when the whole GameStop fiasco happened. So it was a wild experience.
Demetrios:
That is wild. And especially, I feel like you being a growth engineer, you were probably in the eye of the storm.
Peter Dun:
Yeah, absolutely. I think, you know, like, you know, with the whole GameStop thing, at that point, we were the ones who had to kind of take GameStop off of the exchange. And the crazy thing was, we didn't know why this was happening or why we were doing this, and we were probably equally in the dark as. Yeah, like, outside of Robin Hood as well.
Demetrios:
Yeah. So you weren't a member of r slash bets. What is it?
Peter Dun:
Wall street bets.
Demetrios:
Wall street bets. That's it.
Peter Dun:
I see. I think all of us were observers in that community.
Demetrios:
It's pretty fun. That's great. But, yeah, give me, like, the day in, day out as a growth engineer. What kind of stuff were you working on? Yeah, totally.
Peter Dun:
So, as a growth engineer, I was pretty focused on kind of like the onboarding piece of Robinhood. So if you've ever used Robinhood before, when you sign up for it, you basically have to go through this pretty extensive, essentially a form where you're filling out your personal information, your financial information, your Social Security number, and then we all do all these background checks on you, make sure you're a real person, kind of understand your financial investment profile a bit better. And based on that, we'll approve your account, we'll do further review. We'll customize your Robin experience based on the type of investor that you are. So that was kind of the main part of the experience I was working on. That's awesome.
Demetrios:
Okay. And what were some things that you did to make it easier or make it more reliable? That kind of stuff, yeah.
Peter Dun:
So, you know, I think when I was at Robinhood, what I really saw was how important that initial onboarding flow is. At the end of the day, it's kind of like this form that you're filling out. But at the volume that we add, which is millions and millions of users, the conversion rate is actually super impactful on the ultimate performance of the company itself. A 1% difference in the number of people making it through is millions of dollars for us. Beyond the complexity of the flow itself, all the logic that we're doing under the hood, you're running tons of experiments on it. I think over the first few years at Robinhood, we ran, like, 50 plus growth experience on it. We were testing anything you could think of. We were even identify if the users on Android or iOS, and depending on their platform, we adjust the income brackets that we would ask in the questions about what income range you were in.
Peter Dun:
Wow. But basically, based on that, we end up doubling our conversion rate. It was a game changer for the company as a whole. Holy spirit. Yeah. So that's kind of where I kind of saw it was, like, crazy important for us.
Demetrios:
That's so cool. And so I'm guessing you had dashboards of how these different tests that you were working on were affecting the final outcome.
Peter Dun:
Absolutely. Like, the entire kind of, like, onboarding flow was homegrown. Our analytics, our experimentation platform, that was homegrown as well. And, yeah, like, every single a b test that we ran, we kind of saw, like, oh, like, how are these, like, different variants doing? And then, you know, kind of understand, like, you know, what's doing well, what's not doing well.
Demetrios:
Wow. Was there one thing that stands out to you as, like, a huge growth lever that you saw when you implemented it. It made a gigantic, like, outsized returns that you got from that one tweak.
Peter Dun:
Yeah, I'll be honest, I think the. I actually don't think there was, like, one thing that was just, like, game changing, but it was kind of, like, the combination of, like, so many different improvements that we did. Like, every individual improvement maybe was not super game changing, but as a whole, it was, like, really impactful for the company. And I think the crazy thing was, like, you know, like, there's a ton of, like, different stakeholders on the team working on these onboarding flows.
Demetrios:
Right.
Peter Dun:
There's, like, the PM, there's, like, the designers, there's, like, the engineers. Um, and there's a ton of, like, uh, like, conflict, too, between the different teams.
Demetrios:
Right.
Peter Dun:
The PM's are kind of the business owners. Like, they want to be getting these changes out quickly. They. They care a lot about the final outcome here. The engineers actually don't really want to be working on this day in, day out. Like, they'd rather be working on, like, more core business problems than, like, you know, tweaking, like, the ordering of steps or, like, adjusting, like, the types of questions that you're asking.
Demetrios:
That makes 100% sense. The I've actually. When you talk about it was more like a whole lot of singles as opposed to one home run. That makes complete sense, too. I've heard the Doordash platform team talk about it as, like, we're just trying to go for 1% better every day, and we're not really looking for that silver bullet. We just want to each day make one step in the direction that we're trying to go for.
Peter Dun:
Of course, it's one of those things where, like, at the beginning, right, like, there might be kind of more, like, kind of no brainers that you're able to make these changes and you see, like, a more significant conversion bump. But, you know, as you're getting into it, like, there's so many things that will improve the experience that might be a little more non obvious, but, yeah, like, in aggregate, it's actually super impactful.
Demetrios:
So were you just going up to friends and asking them to sign up and watching them as they were signing up?
Peter Dun:
No, I mean, at that point, we had enough traffic where, you know, it'd kind of be more of, like, a numbers thing. Uh, you know, we had, like, thousands of users, like, daily, like, signing up, and, uh, yeah, we. We'd be able to evaluate the metrics from. From that point.
Demetrios:
That's awesome. So give me the inspiration and how Feathery came about. Yeah, totally.
Peter Dun:
So, um, yeah, I think we just kind of talked about, right, like this was the onboarding floor of Robin Hood was essentially this giant form. Um, and it was like really hard to build. Like ton of logic and tweaks always going in. So we do look at the market, evaluate. Are there third party tools that could help us in this building experience? So we looked at form builders, we looked at form builders on the market. And when you think forms, you probably think Google forms or Typeform. And these tools are great for a simple survey or a simple questionnaire. But if you're Robin Hood and you try to build your core kind of onboarding flow, like, these tools aren't really built for that.
Peter Dun:
Just from like a design logic perspective is not really there. So that's where really the inspiration for Federer came from. We are a form builder. That's like a big aspect of our products, but we're designed for companies like Robinhood and enterprises who have like high level needs and styling and logic to be able to actually build these kind of core product experiences.
Demetrios:
Well, especially when you think about Google forms, there's no way that you can have a product that is going to inspire a lot of trust in users. If the way that you onboard onto the product is a Google form, maybe an MVP. I totally love that hustle. But if it's like, I got to pay for this product and I need to go through a Google form to do it, that's not going to inspire trust.
Peter Dun:
I mean, you want to feel like they built it right? Like, you want to feel like it's part of like the product experience. And, um, I mean, actually, a lot of companies do use like a Google form or a type form for kind of like their initial, like prototype onboarding flow. But then, like, at some point, obviously you're going to move off of it and today, like, people are just building this stuff in house.
Demetrios:
Well, also, I think Google forms in type form falls over when it comes to the whole a b testing. You can't really a b test those that well. And if you do, the metrics have got to be a mess to deal with. There's going to be a lot of janky stuff that you're doing in the background to see which one actually helped conversion and how much better was it?
Peter Dun:
Yeah, so I'll probably start just by saying, like, if you're in the Google forms or type forms stage of like MVP'ing things out, you're probably not at a scale where you care too much about the a b testing yet. You probably don't have enough traffic. You're probably not looking to optimize things enough yet where you care about that. But once you hit that next phase of you, an actual sign up, onboarding information collection flow, it needs to be a proper flow. It has to integrate with the right systems, the right logic, the right visuals. That's probably also the point where you're like, all right, how do I really start making this experience really great? Running a b test, that type of stuff.
Demetrios:
And so now tell me how AI came into the mix. Yeah, for sure.
Peter Dun:
So fed three is a product. I'd say, like, our core market is the types of companies who have very complex forms. So Robinhood is actually one of those, because Robinhood is a financially regulated entity. So the onboarding flow is pretty extensive because by law, you're required to do KYC or know your customer to identify the user. Signing up for the platform, which means you have to collect sensitive information from the user. And then it becomes a game of like, how do you collect it in the most frictionless and delightful way, where users aren't dropping off. And this is true of financial services as a whole, whether you're a bank or you're a lending institution, or you're an insurance company, all of these types of companies have very extensive forms where they're onboarding their customers and clients for the purposes of extending a loan or extending an insurance policy or anything else.
Demetrios:
New credit card? Yeah, 100%.
Peter Dun:
We serve a lot of these types of customers. Financial services, healthcare. What we noticed is, with the sheer quality of information that you're collecting from your users, a lot of this information can be, it's great. It's a drop down, or it's like a radio button, easy. But they're also uploading financial statements or a lot of these insurance workflows. This is where the audio piece comes in. They're collected over the phone, or you want to ask the user questions, and the user wants to verbally just give you answers for when they want their insurance coverage to start and stuff. So we were like, oh, wow, there's all this unstructured data that people are already collecting through Feathery.
Peter Dun:
How can we help them leverage this data even more? We realized that we were in the perfect position to take advantage of the recent advances in AI because we had all this information that we were sitting on that people want to understand. That's where we started. Launching are two core AI products. The first one being document intelligence and the second one being voice intelligence to help customers extract just more structured information from their user provided info.
Demetrios:
And so break it down for me. Is it within Feathery you now have these different products that you can plug into your forms?
Peter Dun:
Yeah, totally. It's natively integrated with the products. We have a voice recording component that you can literally just drop into the form experience that you're building. For example, we have a file upload component. When users upload the files, you can set up so that it automatically runs the right extractions, prompts, classification on these uploaded files.
Demetrios:
Right.
Peter Dun:
Like maybe they upload like, their fidelity investment statement. And you're like, I want to parse out like, all the stocks that they have, all the holdings that they have, the prices, the cost basis will automatically do that for them.
Demetrios:
Wow. So cool. And that feels like a game changer, because you now are not limited to just like you were saying, like, yes, no questions, radio buttons or checkboxes.
Peter Dun:
Yeah, it's a whole nother dimension of data collection, I think. And just with the types of companies we're working with, they're very document heavy. When you're getting an insurance policy, you're getting an entire PDF back with information about when does it start? How much are you covered for? A lot of these institutions in finance and healthcare, they're still very this paper, digital document driven to this day. I think that's kind of a big part of where feather resets too, is digital transformation. Yeah. Helping companies digitize and just activate their kind of paper document data better.
Demetrios:
And are you also helping these companies do their A b testing to try and see, hey, how, how much better of a conversion rate can you get now that you're in this position? And it kind of comes out of the box, I imagine, with Feathery.
Peter Dun:
Yeah, I describe Feathery holistically as like, as a whole, we are a form and workflow solution. So there's that form part of it which is the user journey, where you want your users to be high conversion, having a frictionless process, onboarding onto your product experience. And there's the workflow piece, which is less user facing, more internal facing, which is once all this data has been collected, how can the company best activate the status? This means review and improving the submissions, being able to wrap the data into their systems of record, like their salesforce and their erps, and also this AI component which is extracting information from the records, reviewing and sending them off.
Demetrios:
So I bet there's a pretty nasty problem of like data quality when you're ingesting all these PDF's, and there's a ton of tables that you think or you ingest, and then you would imagine that everything goes well, but it's kind of hard to check.
Peter Dun:
Yeah. I mean, that is the problem in document intelligence and voice intelligence, and even in AI, right, which is you're automating all this stuff, and then people want to know, right? Like, what are the, like, are there mistakes in the data that we've collected? Like, what are the mistakes? Like, how can we, like, fix these mistakes? And so that's actually a big value prop of why Feathery has been so popular in these spaces is because we also have kind of like, the schema building piece of it, which is like, you can define your schema in the form, right? Like, I. This particular type of data, the policy type has only four options, and it can only be one of these. And if the user enters something else, then it's invalid. And so data validation of these inputs and letting people create the exact type of data validation rules that they need just enhances the whole experience. In finance, they call them DIGO, which is like. Like a wall advisory firm might collect all this data and try to open an account for their customer, but then the data might just be bad, and then the custodian calls it a NIGO, not in good order. All this information is bad.
Peter Dun:
You have to redo the entire thing. It gets really painful. So data validation is super critical for these people.
Demetrios:
Yeah, it feels like it is top priority just to save time. And that back and forth of. All right, we tried. It didn't work. Can you fill it out again?
Peter Dun:
Yeah, exactly. There's a fine balance between automation and also kind of human the loop review and validation here, man.
Demetrios:
Well, it's fascinating to geek out on that. I feel like the power that you're bringing not only in the. Just the product of having a form that is built, like, purpose built for this problem that you're tackling, but then also the maturity that you're bringing to it when it comes with, like, oh, yeah, we can a b test stuff. We can tell you if you want to have, like, these data validations, or we can bring the unstructured data into the mix, too. Just, it feels like a very cool product. And I. I wonder, you mentioned about integrations and how you want to have different integrations. Once the data comes in, then you got to figure out, all right, now where are we piping all of this data? And there's a lot of unstructured data that you're dealing with.
Demetrios:
And then there, I imagine there's structured data that you're dealing with that also feels like a bit of a hairy problem.
Peter Dun:
Yeah, for sure. Super hairy.
Demetrios:
Yeah.
Peter Dun:
And speaking of unstructured data, just kind of talking on the voice AI side of things a little bit. So this is actually where we started working with Deepgram. It's actually really interesting. So insurance companies, when you're applying for a policy, there's, insurance companies usually fall into two buckets. There's a bucket of kind of low value policies, maybe like renters insurance and stuff. You go online, there's this automated process. You fill in your info, you get a policy, easy. If you're trying to buy life insurance or home insurance, you're talking to a broker, you're getting on the phone with a broker and he's asking all these questions.
Peter Dun:
And then on his side, he's manually entering all this information into a word doc or a spreadsheet. And then he manually sends this information onwards, uh, you know, to the carrier. And then the carrier looks at this data and they're like, what the heck? Like there's all these issues here, right? Um, so that's kind of what's going on today. And so actually what voice AI has allowed us to do is, um, you know, as, as they're talking, like we're auto transcribing and parsing the information, um, into these schematized, uh, forms, uh, and validating the data at the same time. And then the broker doesn't have to worry about like, oh, like I'm taking, like make sure I get everything down and take all these notes while they're talking to the user. After the fact, they go back, they look at what's been filled out, they review it, they're like, looks good. They sent it on the classic looks good to me. Yeah.
Peter Dun:
So, yeah, so, I mean, it's, I think AI has just been a really powerful piece of this whole puzzle, dude.
Demetrios:
So that is a way of looking at forms that I hadn't even have thought about. Because when you think about forms, obviously you're always kind of, or at least I default to me and a computer. But this is a form that needs to be filled out. Usually it's filled out because I'm asking you questions and then you're answering those questions and then I have to manually fill them out on my side. But you're saying, no, we can, with the voice AI piece of it, we can go back and just let that happen on its own automatically and you can figure it out as it's going, I imagine you can just be looking at it and be like, yeah, that's what I said. That's how I said it.
Peter Dun:
Yeah, like, I mean, you're on your side. You had the form, right? You're asking questions. You're like, all right, I covered this point, this point, this point. Oh, I mean, I haven't covered this piece of information yet. Let me go through this one. Um, yeah, so it's kind of like a guide for the broker as well. And the reason why this is, is because, you know, when you're buying high value insurance, you want to be talking to someone. You want to be asking them questions, you want to feel like you're working with someone that you trust.
Peter Dun:
And at that point, um, you know, this kind of, like, just digital interface on your laptop, um, doesn't create the best experience for you. And insurance companies are, you know, obviously always thinking about what's going to get users to convert the best. And oftentimes, it's actually a human, essentially a human form, you know?
Demetrios:
Yeah, yeah, it is a human form. And that is so funny to think about it, because when I think about my own experience, like, yeah, if I'm about to drop a lot of cash to buy something, I don't want to do it through a form on the Internet, because I feel like it is so impersonal. Even though the human form is kind of the same thing. I'm just talking to a human, and then they're doing it on their side. So it's like, let them fill out the form for me.
Peter Dun:
Yeah, I mean, yeah, I mean, it's the same in the hospitality industry and in medicine.
Demetrios:
Right?
Peter Dun:
Like, people need that human touch in a lot of their interactions, even though it could theoretically be digitized. So that's why Feathery is a product. We have that digital component, but the human and the loop piece of things is also super important for a lot of information out in workflows.
Demetrios:
So tell me about where you see things going. What is your vision now, from now, for the next 6-12 months?
Peter Dun:
Yeah, totally. Um, so, obviously, we're doubling down on, like, the product itself, like, working on the user experience, workflow features. Um, the AI piece is a huge part of it. Um, and that's probably gonna be one of our biggest focuses moving forward. Um, to your point about data quality. Right? Like, that's an ongoing battle for us. You know, these financial documents can get insanely complex. Like, 30 page, like, all these massive tables.
Demetrios:
Right?
Peter Dun:
Like, and these finance companies have, like, really high, like really strict requirements around like how good the data has to be before it's useful.
Demetrios:
Right?
Peter Dun:
Like they're looking at like 95%, 99% accuracy rates, which is like a far cry from where document intelligence was like, you know, ten years ago. Back then, you know, they'd evaluate these solutions and, you know, our customers would tell us, they'd be like, yeah, we get like 30% accuracy, tops. And I really feel like that Intelligence is poised to make a major breakthrough with all these recent LLM developments.
Demetrios:
Right.
Peter Dun:
Because we can actually get up to these accuracy levels where, uh, financial services feel comfortable enough not needing people to then go and review every single document. Once you're at 99% accuracy, you're at human level accuracy. So for us, it's really about like just optimizing that experience, keeping accuracy up, keeping the speed up to give them a good experience on the product.
Demetrios:
Yeah, that's fascinating that you mentioned that, because 99% accuracy, you can't just like upload one of these financial documents into chat GPT and get 99% accuracy out of the box. So there's a lot of work that's happening on your end, I imagine, trying to figure out how to get those extra percentage points on the accuracy side of things to make sure that you're, yeah, you're ahead of the curve and you can have that trust.
Peter Dun:
Yeah, yeah. Um, yeah, there's a ton of components to it. Like, I mean, you start with the pre processing, right? Like making sure that, um, the documents, like, you know, sometimes people will just take like a, you know, crappy image of like blurry image of some PDF and upload it. You have to rotate it, crop it, fix the resolution after that. GPT is part of the system that we use, but it's not the only piece that we use. Different systems are really great at different aspects of document recognition. Then we use a hybrid system that we fine tuned to produce that final result for the user.
Demetrios:
Incredible. Well, Peter, this has been fascinating, man. I really am excited for what you're building and I wish you all the best with Feathery.
Peter Dun:
Thanks, man, I appreciate it. I had a really good time on this.
