AI for Everyone: How Gooey.AI Empowers Global Frontline Workers with Low Code Workflows - Ep. 244

0:10Hello, and welcome to the NVIDIA AI podcast. I'm your host, Noah Kravitz. Our guests today were recently featured on the NVIDIA blog for their work in creating Ulanghizi, an AI chatbot that delivers multilingual support to African farmers via WhatsApp. As vital a project as that is, however, GUI AI is much more than a single chatbot. GUI AI is a platform for developing low-code workflows built on private and open-source AI models. Combining ease of use with innovative features like golden Q &As, GUI enables developers to code fast and change the world. Here to tell us the GUI story are the company's founder and CEO, Sean Blagsvet, and founder and chief creative officer, Archna Purcell.

0:52Welcome to you both, and thanks so much for joining the NVIDIA AI podcast. Hello. Hi, thanks, Noah. So there's a lot that I'm looking forward to you getting into about the GUI platform, how it started, all the things it can do, including how you're helping developers combat AI hallucinations, which is a big topic these days. But I'd love it if you can start at the beginning and tell us what GUI AI is and how you got started. Let me take a shot at that. We got started from actually a digital arts project funded by the British Council many moons ago, I'd like to say 2018, 2019, where we applied to create an AI persona that would matchmake creators, activists, designers from across borders of the UK and India.

1:39And we won that award. We built out a prototype. We tested it. It worked beautifully. And, you know, long story short, we managed to get funded a seed fund from Textaz, isn't it right? Yep. Yep. We got into Textaz, which was an excellent program. We took this idea of an AI persona and built an entire communications app around it called Dara.network, meant to service cultural organizations and social impact organizations, enable them to manage their alumni and keep in touch with each other easily. The first AI persona we built, also confusingly called Dara, we were feeling lonely and wanted some friends.

2:17And we thought, wouldn't it be great to invite non-tech folks, writers, playwrights, authors, poets, could we have them come in and craft their own AI personas from scratch? And this is right in the middle of COVID. So all those are out of work. They're out of work. They're isolated. isolated yep yep yeah yeah and so we invited 23 folks from across the uk the us india sri lanka even and met constantly literally every week and ended up developing pretty much a underlying architecture that enabled them to build out what one might call a hurrying test testing video box right? So our co-founder, Dave, was hanging out in Discord forums with, what's his name, Brockman, the president of OpenAI until I think it was recently.

3:08Chris Brockman, yeah, yeah. Yeah, it was like five years ago. And then so we had really early access to the GPT APIs. And we had already built Dara as an asynchronous video messaging platform, kind of like, you know, Discord plus LinkedIn with a little bit of Mark Polo in there. And so the thought was, well, what if we took, you know, it's kind of a wild idea. Like, what if we took the video messages that people sent, hook that up with Google's, you know, speech recognition, fed that to a long form script, right? These playwrights and authors were putting together. And then, you know, we had this thought of like, what is it if it's$1.50 to$2 per API call?

3:47What could we get, right? And then so we basically had this script that they were crafting and they were writing together. And then we had, you know, the deepfakes APIs were just beginning to come out and there was speech. So we're like, well, we can take what the bot says back, right? That DaVinci output and then take that little text, put it to a text-to-speech engine, put it into a lip sync piece. And then boom, you've got these Turing test passing characters. We call them the rad bots. And those were the rad bots. They were awesome. They still are kind of awesome. Yeah, they were crazy. And like we had a little thing.

4:23This is a public family podcast, but, you know, we had, you know, the rad bot say. That was the Wild West of LLMs back in 2021. Yeah. Yeah. And these bots really spoke their mind. And they just spoke their mind. They represented issues that the writers brought in that they felt were not represented well enough. or spoke on behalf of communities that were underrepresented, kind of in a bid to reduce the algo bias that the group was feeling quite intense. You know, you're in, the bots are done, they're happily chatting away, people are excited. And like little kids have talked to them a thousand times, like one kid, 1200 times, right?

5:04And these were presented at, you know, the... Yeah, India Art Fair, the, you know, Be Fantastic, the British Council hosted event, to the India Literature Festival. They went to it quite a bit. The point was that we had built out this underlying architecture and orchestration platform pretty much that enabled us to kind of plug and play all the incumbent new technologies that were emerging in AI. And that was a moment, I think, in this very room, actually, where John and I were like, okay, we have the messaging platform, but what can we do more? And that was pretty much the start of GUI. We felt, hey, if we can take artists and writers with us on a journey, why not open it up into a wider world and kind of really take this, you know, this mission we had begun to feel quite deeply by then of like democratizing AI and really allowing people to play with it, even if they didn't know how to code.

6:02I'm not a coder. So, you know, I feel that part. I want to ask you quickly about your backgrounds. I was wondering kind of both technically and at the risk of making, you know, being corny. Aside from the fact that enabling everyone, artists, creatives, non-coders to use the technologies, aside from the fact that that's just an awesome thing to do. Why was that your philosophical? Why the hell would you do such a thing? Well, my background is actually in art and design. I studied painting and animation film design back in the day and quite wonderfully met Sean. he had landed into India to set up the Microsoft Research Lab.

6:39This was 2004, 2005. So I was number three and she was number five. In that amazing organization. We're going to track down number four, get them on the pod. Yeah, so that's how we met. And I don't have a coding background, but I did get to hang out at Microsoft Research, get access to some of the brightest minds in the world and the papers that were coming out of there. And we were the, I'd say the first and maybe the only prototype. What is it? Advanced prototyping. Mike Heidel was head of product management and advanced prototyping. And so we did a lot of work. You know, this is like 2004 to 2010 around, you know, how do we build interfaces that reach everybody?

7:19How do we build interfaces for folks where literacy is an issue? You know, I have the patent on, you know, machine translation and instant messaging from 2005. Oh, yeah. This is a space that we have been in for a long time, running big models on NVIDIA hardware, trying to understand what everybody has to say through chat interfaces. And frankly, India was ahead that skipped email, right? And everybody kind of went directly into SMS-based interfaces, into WhatsApp in the years ahead. So this is a space that I would say we've been experimenting in literally for 20 years in terms of like, how do you build tools and interactive AI based interfaces that work for everybody, including people that don't read or write in English very well.

8:01And then after that, Archana went off and set up an organization called Jaga and then another called Be Fantastic with co-founders Freeman and Kamiya. And the idea was really, how do we take artist and creative practice and technology practice? How can we bring practitioners from these different fields together and enable them to have conversations and service kind of some of the big pressing issues of the time? So soon a lot of activists and all that started. We started JAGA in 2009, Freeman and I, and Be Fantastic kicked off as a public arts festival looking at arts and technology, particularly.

8:40towards climate change and UNSDG kind of issues. And then meanwhile, I was off running a company called BabaJob, which was the largest informal sector-focused job site in India. So how do we take drivers and cooks and maids and basically using the phone and IVR and SMS and multilingual interfaces, basically hook them up with better jobs. You have a pretty deep phone interface. Yeah, she was the voice. But, you know, we did 50 ,000 phone calls a day. Our telephone bills were astronomical. We had 9 million users. reprocessed a million applications a month. That was a big initiative. That was 11 years of my life while she was running Jog and Peril.

9:18So there we got there. Yep. So you've got the credentials, you've got the chops. And I think before I interrupted you, Archana, you were about to say, and so we wound up with this platform that we called GUI. Yes, we did. And I love the name. It's a little take on, you know, graphic user interfaces. We talked about old times when GUI was a thing. Also the connective tissue. And the GUI laugh, you know. Yeah, so we kind of pivoted pretty much overnight, quite literally. We took our team over to the idea. They really liked it as well. And pretty soon we had GUI, which is pretty much what it is today.

9:54Well, and then, you know, it was sort of founded on this premise that almost started out as a joke, but feels less and less like a joke every day, which is that all of us are just going to be AI prompt writers and API stitchers, right? In terms of like the job of the future that everyone will do seem to be that. And then, you know, this thought being, well, if that's the world, what kind of things do you need? Like, what would be the JS or JS Fiddle equivalent of that? As in how, when I make something, do you get to view source on that and understand what I'm doing so that I can learn by building on top of what you've done?

10:26Right. Kind of like also from the, you know, publishing, I mean, yeah. Yeah. The research world of citations and public papers, which goes back to the beginnings of the Enlightenment, right? And so, and the open source movements as well. We wanted to say, what would that mean for these new higher level abstractions of, hey, you got a little bit of LLM prompts and you want to connect over to hear this other API and you want to connect to some other communication platforms? How do you extract that up to sort of allow a whole new generation and non-coders to basically play? And efficiently, too, right?

10:59I mean, the idea being that it's a one-stop spot, right? You can try all kinds of different AI technologies and tools without subscribing individually to any one of them. Yes, it's, yeah. And so that's because of this part of, we saw a huge amount of innovation coming from many sectors, right? The open source ecosystem was clearly making new incredible models every day. Right. Not just around LMs, but around animation and image creation and, you know, text to speech, right? And we saw, given our work with Radbots, that when you allow people who are creative and empowered to put those together in novel ways, you get this magic.

11:38That's where you get the magic. Thousands of interactions, which is definitely bigger than the sum of its parts. Yes. And so we wanted very specifically to say, great, when OpenAI or Google or the open source community comes out with some new feature, we should constantly allow the ecosystem to get better. So with this whole thought of we should abstract on top so that every component is essentially hot swappable and evaluatable, which is where your golden questions things come in. Yes. That you can basically say, hey, OpenAI 4.0 or, you know, some O1 piece came out. Is it better, cheaper, faster for me?

12:13And then, you know, given our impact piece of like going forward, like, is it also what's the carbon usage of each one of them? Okay. Right. And how do we make that clear to those people that are buying and using those APIs as well as sort of the fourth important factor for any chain or workflow that you're going to put together? So many questions. Maybe I'll ask you, and this is sort of a bad question for an audio-only format, but that's what we do here, so we're going to do it. A new user comes to GUI. Yes. Someone who perhaps, you know, understands what an LLM is, how technology works. They know what an API is.

12:49Maybe they copy and pasted some code once or twice, you know. How do they get started? Is it drag and drop? Is it writing things into, you know, sort of a chatbot, text-only interface? How does the platform actually work for the user? So at a high level, you know, this is something we took from BobaJob. We did pretty good at SEO. And one of the reasons that we do well at SEO is we try to lower friction. So when you come in, if you Google like AI document search or, you know, agricultural bot or AI animation, you can come in there and you can see the prompts and the output directly. And you can go into examples and you can see a bunch of other ones that we hopefully gets better and better and are relevant for your field is sort of a UGC model.

13:33And every single one of them, you can just say, great, I like that. I'm going to tweak it, right? Right. So you grab a pre-existing, yep. I'm going to change what this model is. I'm going to, right. And so we're always having this kind of, we're definitely inspired by like replicate, right? Which is definitely, you know, this idea of like, what are the inputs somebody else used? What are the outputs? But to do so in a way that we're chaining together many of these different components, just basically see something awesome. So that's kind of it. And it's a drop-in more kind of pull-down menu. Yes.

14:00Because the idea there is transparency. Like a lot of other sites, I would argue they hide the prompt that's actually making the magic happen or they hide the model that's making the magic happen. And so for us, we have a belief of like, no, the model that you're using for today versus tomorrow and all of the prompts and everything else. It's all changing black box here. It's all digestible and viewable and inspectable all the way down. We kind of call these recipes for a long time. Yeah, sure. Yeah, sure. Yes. Really, like these are the ingredients. This is how we made this mix. And, you know, you can make your starting from there.

14:37Yeah. So everything is forkable. And then again, just like JS Fiddle, you make one change. That's a new URL. You can share it with your friend. You know, literally next week we'll come out with workspaces. So you can work on these things collaboratively with version histories. So you can say, hey, I have a static endpoint of like my cool co-pilot. We can work on that together. And then you can do things like hook that up directly inside the platform to things like WhatsApp or Slack or Facebook. And that's actually, I feel like, an underestimated part of getting these things to work on the communication tools that real people use is way harder than you can do.

15:13There's a lot of friction there that we try to take away. Right. And I wanted to mention, you know, and I don't know, this is back in the day, I covered the mobile phone industry. And I don't know, maybe we have a great audience, so they probably know. But, you know, for sort of a U.S.-centric point of view, people don't necessarily understand that in so many parts of the world, your phone is your computer, period. And people are sharing phones or, you know, getting a phone to use for a day or that kind of thing. But it all happens on the phone. There's no laptop. There's no desktop workstation, all that stuff.

15:43And so when I was, you know, researching, prepping for the taping and reading about how Ulangizi, the farmer chatbot, went through WhatsApp, you know, it was like, oh, cool. And I was like, well, of course it does, because that's how that's how people work. So, you know, maybe to get back to what you were saying, Sean, about putting these tools into the communication platforms. What were some of the hurdles, some of the challenges, maybe some of the pleasant surprises in working on that? No, there's a ton. Right. And so we've got a talk that's on our site. called like the stuff that goes wrong, right?

16:17Which is basically like, so, you know, right after we began, again, our old friend and director of the lab, Anandana, you know, said, hey, Rickon, who runs Digital Green, is working. And the whole space is like, you know, bots for development is going to be a thing, which is basically getting this part of, if we want to convince every farmer in the world to basically change their livelihood and the crops they grow, because climate change is necessitating that over the next decade. We've got to convince all of them to change literally the source of their income. That is a hard challenge. Every government in the world has this challenge for the, I don't know, several billion people on Earth that are farmers.

16:58So there's this piece that he recognized like, okay, bots are going to be a thing. One should get together. And so what we did there was to say, hey, you know, in the case of Digital Green, They had an incredible library of thousands of videos, basically of one farmer recording how they do a technique better that was then shown to other nearby farmers. Farming best practices. Farming best practices. There's also, you can think of it as like all of the fact of every question that everybody's asked and the ag space that goes to the government. And then a bunch of like local knowledge in the form of like Google Docs of like what people should do coming from local on the ground NGOs.

17:37And so what we did is to say, hey, we've built this extensible platform. We can have ragbots. We know how to do speech recognition well. We're running private keys to all of the best services. Plus, we have our own A100 infrastructure and GPU orchestration. So we can run any public model too. So then we can say, great, we can take all those videos, which are not in English, right? Transcribe them. Basically use a bunch of DP2.4 scripts to create synthetic data around them so that it's not just this transcript, script, but it's also like, what is the question that a practitioner might actually ask?

18:09And what's the answer here? And then use all of that to basically shove that into a big vector DB, right? And then say, okay, we then hook that up on WhatsApp. And then you put in translation APIs and speech recognition APIs in front of that. And then boom, you suddenly have something that works in multiple languages in multiple countries using locally referenced content with citations back that can speak any language that is actually useful. to folks on the ground who are small shareholder farmers. That was what we demoed at the UN with Rickon in April, 2023 at their General Assembly science panel, right?

18:43And so you now look across the world, the bots are a thing. And I'm not saying like, obviously we're the only people involved in this kind of transition. But the thing that I think for us was exciting is a bunch of people in the private sector also noticed. And they said, hey, if you're looking at how do I make frontline workers productive, people that need to fix your AC or do plumbing, they have the same issues of like, I need to aggregate every manual of every AC sold in America, plus all of the training videos around them, plus ask any hard question in order for me to do my job. And oh, by the way, all the master technicians in that field retired with COVID, right?

19:22And so there's none left, right? And so, but the technology that you need to make that happen is actually the same. And hence, you know, You'll see us, we talk a lot about frontline worker productivity because I think we do this really well by essentially aggregating all of these different parts. That was the long answer. Yeah. One of the things that you mentioned a few times is languages. And, you know, a lot of the models, I mean, English, for better or for worse, is taking over, spreading, ubiquitous, et cetera, right? And a lot of the models trained on English, you're working with all kinds of languages, including, from my understanding, tons of local dialects and, you know, the kinds of things that the models aren't necessarily trained on.

20:07Right. Tackling that, right. Talking about translations and all that kind of stuff. Are you also working with, you know, training foundational models in these languages or is it just a better way to tackle it by doing and I may have this wrong, so please correct me. But doing what I think I understood as translating back to English and then using that to work with the LLMs. Again, it goes back to the sort of core philosophy of GUI that we always want to be the superset of everything else out there. I personally think as a small startup by small, I mean, under a billion dollars in funding. It is fool's errand to try to train any foundation models.

20:47Yeah. Right. Because every six months you're going to be outclassed. And so I'm going to leave that to the people that can put a hundred billion or more into it. And yet every single day I want to know, does that work better for my use case? Sure. And we take this very use case specific evaluation methodology, which is this golden and questions and then apply that to, hey, I have 50 farmers outside of Putna in India speaking this particular dialect of Bhojpuri, right? Here's the questions that they ask. Here is the expert translation or transcription into Bhojpuri. Here's the expert translation of that question.

21:23That is my golden set. And then what we allow you to do is to say, I'm going to run this essentially custom-made evaluation framework across every model and every combination of those things so that this week I can tell you how the Facebook MMS large model works actually better than Google's USM, which may suddenly work better than, you know, GPT-4-0 audio, right? And to basically allow organizations to evaluate which of the current state-of-the-art models, and in particular, the combinations of those work best for their use case. So we have an evaluation level, not the training level. Right. Is that a hands-on user thing, figuring out which model, which combinations to use?

22:08Or is that something the platform does for the users? That itself is another workflow. So gui.ai slash bulk, right? You can upload your own golden data set. And then you can then say, great, I want to do a, again, you can see all of the work that we've done for other organizations. And then you can just sort of say, great, this is how they did it. I can copy, not copy, I can just fork their recipe on the website. And the advantage there is you don't have to run the DevOps to run all of those new state-of-the-art models. No, absolutely. I'm speaking with Sean Blagsfett and Archna Prasad. They are the co-founders of GUI AI, a low-code, change the world, literally change the world.

22:48A lot of people say that, but I think y 'all are doing it. Change the world platform for using AI models for all kinds of things, but we're talking particularly about frontline workers, be it an HVAC technician or a farmer in a rural community in Africa. Sean, you mentioned, I teased this at the beginning, you talked a little bit now about the golden sets and the golden Q &As. So I want to ask you about that and about issues around hallucinations. It's one thing if, you know, I'm using a chatbot to help me in my writing work and it hallucinates and I can sort of read it. It's another thing if a farmer or anybody else is asking a chatbot for best practices for their livelihood.

23:27Hallucinations, literally, you know, life or death there. How do you deal with that? So there's a variety of techniques, right, that I'd say out there. Yeah. You should be suspicious of any time anybody says we're 100 % hallucination free in general. So there's the rag pattern, which says, hey, I will search your documents or video or whatever you put in there and I'll only return. Well, then you get back those snippets and then you ask the LLM to summarize it. The risk of hallucination there goes down, right? Because you said, hey, I'm summarizing some simple paragraphs. That's probably OK, honestly, for things like farming.

24:03It may not be OK for things like health care, because the other thing that happens often in our pipelines is you take that, you know, kind of summarization and then you do a translation and that translation, you know, for English to Spanish. Great. We're not going to probably have a problem. But English to Swahili, English to Kikua, you're like, I don't trust that. So with that other techniques that we see out there, where if you really want to do hallucination free, then what you do is you sort of translate the user's query into a vector search of which question that's already in your databank, whose answer has already been approved by, say, a doctor.

24:41Does your question most align to? And then the information you give back is not the answer to the user's question. It's, hey, here's a related question that I think is very semantically similar to your question with a doctor approved answer. And then you use essentially your analytics, right, to say, hey, how often and how far away is the user's query to the question bank that I have? And then, you know, I can then go get more questions that have verified answers from doctors and make that bank bigger and bigger and bigger over time. And that's how you actually get hallucination free because it's a search.

25:17So that golden set is the vetted questions and answers that you're then searching for to see. Users can't see this. Sean made a face and looked up, so I stopped. Oh, yeah. So those are two different things. Like what I was talking about is what is the knowledge base in kind of a rag pattern? Yes. Golden answers is really the use case specific evaluation frame. Okay, okay. And so you can think of it as most LLMs look at like the MMLU as the benchmark that they should be rated against, which asks a bunch of multiple choice questions for graduate students and things like organic chemistry. That doesn't tell you how to fix an AC.