Hello and welcome to Intel on AI, I'm your host, Ryan Carson, and I appreciate you all joining us today. I am really looking forward to interviewing Sean Phan today. He is the co-founder of Pixel ML. In June 2023, he led the launch of Quick QR Art, an AI powered QR code art generator.
It has since attracted over 1.1 million creators globally. Pixel MLs and innovative work has been recognized by programs such as Intel Lift Off, which we're a big fan of here at Intel and Industry partners like MMI Global. Sean is currently focused on developing agenticflow.ai and AI driven content supply chain solution designed to streamline enterprise marketing processes aiming to help companies create and manage content more efficiently while ensuring compliance. Great to have you on the show today. Where are you calling in today from, by the way? I'm so, at the moment I'm in Southeast Asia.
-Ho Chi Minh City, Vietnam. -Awesome. Good to have you here. So for listeners that are new to AI agentic workflows, can you give a real world example that demonstrates how they work, what they are and more? Yeah. So a theme for this one. I would take an example from professionals and drill from Deep learning and AI.
And I'll say the code for it, of course. Yeah, right? So he was saying that ChatGPT would have been done an amazing job, right? So it's like you are writing a piece of content saying it is like you just as is ChatGPT saying, okay, you write it for me and just don't stop just right from end to end is if I received that task, I cannot do. But amazingly ChatGPT can do. But it just one shot prompting and it get that but the level of coherence and everything is not that good because it's not the way that humans normally do.
Because when you have a piece of content, you are not just splayed out and writing everything you need to brainstorm the key idea, the keyword that you would like to target, and then you do the outline and then you write a first paragraph and then you review with the outline and all of that. So actually this is a like processing pull is actually something that we can mimic that with LLM and that's what we call agentic workflow. And in Andrew's Research he was saying that even with like GPT 3.5, which is very old model with agent Dataflow, it can offer the higher performance even so surpassed the 4.0 or maybe 4.0 at the moment. So the idea behind that.
So instead of like one shot generation and getting the result directly, it will give the AI the time to sell reflection based on the input or instruction that you put in and then based on the research that it is generated. And then if we are polishing and maybe twisting a bit, before throw it to you and this process, having done it internally during the thinking process and that not only applying for for blog post but also apply for any type of content generation the AI by like image generation, video generation or code, all kinds of that can be applied in the same process. that's what's called Agent AI workflow.
And that's why I came up with the name of agentic flow agentic flow mean -agentic AI workflow, that's it. -Yeah I love it. It it's interesting. I was talking to someone and they were explaining that what an LLM LLM doing is effectively it's as if you said, you know, write a huge paper about the history of the United States and you're not allowed to press backspace and you're not allowed to stop. And then seeing what happens. Right?
And that's effectively what LLMs are doing when they generate the next token. And yet, like you said, it's amazing. But with agentic workflows, you're able to say, okay, create an outline first and then think about it and then refine that outline.
Okay, Write the first paragraph and then refine that. And it's like you said, actually the way humans work, which is why it's much more effective. So it's going to be fun to kind of dig into agentic workflows as we talk more. So what is the most popular AI Agent Workflow that your customers utilize right now? What are you seeing they absolutely love in its most popular one. Yeah, so we have quite a lot of pre P workflow in our customers.
So a self building damask as well because we provide no call workflow builder for them to customize and this field of business process. But what I notice is that there are two not one workflow that's really picked up. So one is about the lead contact enrichment.
So just imagine that you have a company, you have your lead coming in from your website or blog posts or something. They have email and name, but when you would like to reach out to them, then you don't have the information background about them to craft the personal online messages that resonate to the customer and then one come up and easier to close. So there's a lot of our customer that using our agentic workflow to lead generation.
So based on their email that we find, their LinkedIn, we scrape their profile, what are the possibilities at least share, and then we summarize and this building through a sort sentence and you know that what are the customers, what is his reason activities, And what is they might be interested on? And sometimes we are also having customers that using LLM to craft up online messages based on that information so that they can just copy and paste and send to customer based on that. And I imagine you could also do lead scoring in the same manner. Yeah. We've got this cold lead come in. Yeah, we have their name and their email that's actually go out and try to figure out, you know, the quality of that lead.
I can imagine an agentic workflow doing something like that as well. Yes, correct. Absolutely. So I love that you describe the company as content supply chain solution. So can you tell us, you know, the kind of problems that you see companies having with this content supply chain and then how agentic workflows start to fix that? Yeah.
So did this happen as a my background? So I do the supply chains for more than half of their and all what we do is optimizing the efficiency and when is coming to the content creation. I saw that that process is very fragmented and also to offset that created in that of the step is not utilize it. So make an example. So we are having our customer that doing is very take a pharmacist company and they are issuing the medical post or email sending to that customer, which is like doctor or hospital and because they have regulated industry. So they cannot just write something and just post, right? they have to go to all channels up like approval, a fact check on this one area and they can spend like months in order to deliver that.
So their problem is that they may have a lot of data. They have a lot of asset or they have the old article or maybe a critical record, but their creative who are writing the blog post, maybe have different level. This guy is writing really good. He do the citation correctly, but these are not a guy is not that good. So the content he is creating is not balanced.
So we can help them by curating and monitoring all of the asset that organization already produced. So those are like footage or maybe PDF file, may call record and so on. We can retrieve them. We give our editor for the creator. They can writing.
And then we can retreat, which is which you just use an admin RAG workflow now retreat that recommendation we give the citation, where is it coming from and then they can be more confident that the content that they are producing is actually true. And then after that, yeah, what they do is that they need to do in some manner, right? They need to have some feature ready made And we also have for them to do creating that as well. Either it can come from their like asset library that own their own company having or even like is coming from the start.
Like I stop and so on. And after that one is it need to be go to approval and either way and right now these are probably email like you send to someone, someone take it like few data we respond back and that's can take a week until that you can have some feedback to like either you complain or not. So we view and we bring them all into one platform so that they can easily working together on one single platform and one more is about a regulated industry. So you can not just post everything you want, you need to comply to a marketing regulation policy of that country and so on. So instead of waiting until the very end, you can actually check it every time that you submit into the system.
Then the system we call is like viewing compliance agent. They would check it and then we tell you, okay, this part maybe not really compliant. You rewrite it. So it fix, and this is going to be much easier to fix when is still at the beginning rather than is it the long chain of everyone. And then you fixes again and they take on that a month and so on. So that is what we are trying to do.
with agentic flows. Wow, I could imagine that. I can imagine a real world example that where you have a pharma company that wants to write a post about one of their products and like you said, it's highly regulated.
You can't make certain claims and you have to be careful about the imagery and and all sorts of things. So you could actually have one agent craft the blog post from a, you know, say there was a press release about the product. All right. We want to write a blog post about that. We want to insert assets that are approved and then an agent, you know, which is essentially called a compliance agent, then checks it, make sure everything okay, fixes it, and then sends it on to be made into also a YouTube short and also an Instagram story that's all automated across agents, which is really exciting.
And I think we'll start to see companies really benefit from tools like yours. So moving on to the future. So let's get out our crystal ball and imagine what the future is going to look like. How do you think people will be interacting with agentic workflows in the future? And I would say because the barrier to entry is right now of lower and lower and more and so getting better and cheaper as well. Just yesterday, Open AI is just dropping their latest model 50% off on the price.
And actually, it isn't going to be amazing for us because we have benefited a lot from that advanced economy. So this need the capability of agentic flow is going to be higher. We can handle more types of tasks that can copilot with the human operator so that they can do more with less time to focus. So these mean some of the ties of like repetitive tasks like policing, possible assessing or maybe checking fact checks or research can be done by the agent helping the operator and is going to be majority of the content that we consume on the Internet is going to be either agent asset to create or maybe fully automated creation or repurposing by the AI that they will based on the blog post that you already have, and they just transform that into another media.
Four part that you can post and is still compliant and it's based on the fact that you already create with the first purpose So it is fine and so on, by the way. Yeah, it seems like they're the first hurdle of getting an agent set up, right? You know, getting the workflow correct, figuring it out, testing it, making it work. It obviously is worth the effort because you might save a thousand times that amount or a million times that amount of time. But how are we going to get over this hurdle that humans, you know, put in front of themselves are just I don't want to do the work to set up an agent.
I don't want to figure it out. I just want to, you know, get the work done today. So are we going to are we going to surpass that or is that always going to be a big hurdle? Yes. So I think you are just spot on. So the learning curve to building a workflow is high and not many people that are willing or having even ability to catch up with that fast. So we need to lower that barrier to entry.
So the first part is we already have the no fly where you can treat, but that is not the first entry point. The first entry point would be your SRP, your process. You already have that process already.
We just copy and paste and then we have the agent workflow Builder going to leap that SRP and then transform. This is like a Lego with buidling block line We have like... 200 building blocks that doing different things like Google Sheet Connector or maybe like Google Drive connector, you may create email, create video or a text. So those are building blocks and like Lego blocks. And then our workflow, assist agent, is going to try to come up with a workflow that's going to solve your task instead of like you come up with yourself. So it's not like the 0 to 1 is going to be done for you.
And then after that you can click and and run it if it's work or not working. And if it's not working, that just edit it all. Maybe tell the agent, will say: Hey, I don't like this one. This happened with you. You do it wrong, you do it again. So what it does that is going to repeat the process through the thinking. And I maybe swapping the Lego block from this one to the other one.
And then over time I think it's going to reach to subtle test go that you can access to use that in order to solve some like, like low value added task or repetitive tasks that you spend a lot of time but not much value comparing to what are you doing. So that is the what investment. And in one company we don't really need everyone to build workflow. We just need a subject master expert to build workflow. Do you think we'll see a specialization for a job role in building agentic workflows, or do you think it'll just be so easy to use natural language that really anyone who is using, you know, tools like Word or Excel will be able to to deploy their own agents? I would say we are heading to the world that everyone can, just prompting to modify the workflow and use. But we actually seeing at the moment that AI agents, they are already start to using these kind of workflow they are like to deployment to a some of the non-tech people because actually you just pay for AI agency and they come, they set up for you and it's done for you. Right?
And they are getting the advertise and this is like the new economy for those AI agencies who just step into the enterprise and having them, it's like SAP, right? So the agents make SAP for you. But in AI workflow. Gosh, that's interesting business opportunity, right? We probably will see the massive number of AI agencies that go in and install AI processes for you and that'll be exciting to see, actually. And I could I could see that being necessary because a lot of enterprises have very strict privacy concerns or data protection policies. And I know that's something that Intel we're really excited about, the fact that we actually make the compute and we can control the privacy and the trackability of it.
So we've got a couple more questions that I love to chit chat about. So could you share a success story or a case study that really highlights the impact of AI agents for one of your client's businesses? Yeah. Yeah. So one of our client is MMA Global is a nonprofit
organization I stand for Mobile Marketing Association is have like 800 companies with big brand that you know about in their organization. So they are hosting the event to foster the latest trending in marketing and so on. So what we having done is that in the events, right, you are joining the event and like 3000 people that joining the event and you would like to amplify an event and there's no better way than your audience that 3000 people is going to be hearing through their social media.
And then I'll tell you the saying that, okay, I'm joining this event, the MMA Innovate 2024, so join me as well. So in order to produce done right now they are just sending email and they say to people like some source Kit but this must this is not personal like but what we are providing to them is that you can have an agent that are going to take the user photo and then it will be bring them into a social frame. They'll say, I'm Sean I'm the co funder of Pixel ML and I'm going to join MMA 2024 and then I'll even generate a caption for me and then the audience the 3000 audience can generate the social frame for themselves and they can post on social media and amplify your event and your impact with now, like your designer had to design for 3000 people. So not really impossible task, right? If you are just doing that design for everyone, but it would be in any when you set up an agent but do it for you and you send that link to your audience and they can do it themselves. And that is one of the obvious use case that we having.
Wow, that's really cool. That kind of goes along with a framework someone told me of. Imagine that you had infinite time. What would you do with that infinite time? And that's usually a perfect use case for an agent because you can deploy, you know, thousands or millions of agents and effectively create infinite time for yourself. I also liked how you said you look at the scale of high impact and low risk, you know, items that are that score high on that. That's an obvious low hanging fruit for agentic workflow.
So let's again get out our crystal ball and think about the end of 2025. What do you think agentic workflows will look like at the end of 2025? So for the agentic flow of why now we are doing the B2B side, So we are trying to be present on big marketplace like AWS Marketplace and Azure. So because we would like to target to an enterprise level customer as they are value to privacy and also they are also experiencing the pain of like having a lot of data, a lot of asset, but cannot utilize that. So they will be in need for a solution that can one click deployment and tie your custom agentic workflow. Now see on your data center for your employees to do working with your data. So those are like the low hanging fruit and also the case study that we are targeting at the moment.
And I'm also talking to to Intel Cloud team. I think you are also preparing for something similar and I'm happy to be one of the first to be put on that marketplace. And I hope that my solution can run on data center that powered by Intel AI, because right now we already partnership with Intel for a while, right? Doing the Intel liftoff like last man in the May innovate 2024. We did an amazing App moment that's going to be the partnership between Intel based and now and one of the number one company in Vietnam. What is it like eBay for Vietnam? This it like eBay for Vietnam. So Tree company we provide is like something AI experience and we get it like a number one interactive.
Our word of the day for that one. As I have had a video I sent you, Eugene, and you can say this is really exciting. Yeah, on that on that front, you are a part of our Intel Liftoff program, which really excited about. What is Intel liftoff provide you and why did you join? Well, I will say I was really impressive about the way that you are organizing the accelerator. I have been joining different accelerator and so on, But this is truly amazing because you give the compute and you give an engineer that doing that two times a week.
Code with us to solve the problem that we are serving. And even for the event that we are saying that and need a Gaudi 2 you know that you like bring that experience. And I was just so out of the group and say hey, can do you think the interesting just sponsoring an event and you said yes and into like what are we going to have we set it up I have a Gaudi 2 and that I said that event we when the first are innovative are what are the MLA global 2024 in the innovate event. I think that's it's an amazing collaboration in support and that that's just amazed me and I'm looking for like the marketplace place and also everything else that you are having layout and yeah, I was really strongly believe that it's going to be a bright future ahead for with the collaboration of it. Why we're so glad that you're on Liftoff and it is pretty exciting. We can offer you actual compute and then compute engineering to help you support you and ideally, you know, see you liftoff.
So wind it down there. I want to encourage everybody who's listening to check out agenticflow.ai and play around with agents I know you can try it for free. Actually played around with it this morning.
I was impressed. Tell us a little bit more about where people can go to follow you on social media or where they should check out more about agentic flow. Will say, my Twitter handle #SeanP_AI you can find me on Twitter or LinkedIn with the same handle. Yeah so I'm mostly active on Twitter and LinkedIn, so you can find me there. Awesome.
Well Sean, thank you so much for your time and thanks for helping us understand where agentic workflows are going to go and where agents are going to go and content supply chain solutions with AI. I really appreciate your time and I'm sure we'll catch up soon. Take care. Thanks, Ryan. I appreciate for the opportunity. Take care.
2024-09-09