The Secret Weapon for Personalized AI Coaching?

We're doing it live. We haven't had these conversations. I got to build an AI health coach or AI longevity coach that's hyper personalized. How exactly do I do that? How do I make sure it doesn't go off the rails and provides relevant recommendations that a user will actually adhere to.

So the first thing you can see here is that in the dataset configuration I can filter it by a user cohort. We have some examples here, like fitness enthusiasts. And it'll basically save, you know, store conversations. I don't think I pressed this big red button yet. And this is actually the whole point of this page is that it's going to run an evaluation. And this uses Deepaval's evaluation toolchain to talk to OpenAI. The coach is going to say, I want to learn about this user's habits or like sleep, diet, exercise, for instance. It runs through the evaluation loop and it gives kind of a discord about how well this is doing. You're like, you're building these interfaces, you're creating a data set of what chats are good, what chats are bad, what's wrong with them.

Error analysis is the single most valuable activity in AI development and consistently the highest ROI activity. If we don't understand ourselves, how can AI understand us? The feature of AI alignment is evidently. Hey everybody. Welcome to the ABCs for building the Future podcast. I am Robert. I am joined with my co host and co founder on epistemic me, Jonathan. And today we have a build in public update. We're going to be going over a core feature. We've identified and articulate its value prop and demo it. That's called the SDK Playground targeting at our hypothesized ICP of a developer tasked with building AI coaches and in particular hyper personalized health and longevity AI coaches. In addition, we did a revamp of our landing page slash website where we've totally threw away the whole design previously and we went for more of a futuristic design with a new color palette inspired by a lot of user interviews and learnings that we've had in the past few months. So we're going to actually go through it. It's already live, but we're literally going to go through it and line by line do some content scrubbing to make sure marketing meets development well. So awesome. Jonathan, do you want to give a quick like word or two on what is the SDK playground for the audience to. To be able to understand it? And then do you want to take me through some of the updates? I Know, in Slack I was seeing a lot of your daily standup updates seem pretty meaty. So I'm very excited to hear this.

Yeah, it's the coordination of a few different pieces and that's, I think that's why the Slack updates are like a working plan for me to implement as I go and remember what the next most important thing to do is. But the FD playground is just a tool that introduces somebody to the epistemic ME SDK and you can basically see in the form of conversational interface between a coach and a set and a user, one user at a time, that you can model different users and look at a conversational interface and then you can inspect like you can see how our API works. The SDK works with the API to extract beliefs during the conversation and how you can use the create and update dialectic techniques to manage this dialectic between a coach and a user. Ultimately it's a, it's also a training ground. So you're training the coach to like act in a way they can understand the user better incrementally. It meets a learning objective and yeah, so we can just jump into it.

Let's jump into it. And for the audience, while Jonathan's pulling up, this is a true build in public update where I haven't seen this and I'll be seeing it at the same time that you're seeing it and then on the website update as well. This is, we're doing it live, we haven't had these conversations, so super built in public, very raw.

Nice.

Cool.

So this is the dashboard for the SDK playground and it's very much a work in progress, but let me walk you through it. So I think in previous iterations when I built that, I've worked a little bit from a bottom up perspective in terms of, okay, what's the best way to demonstrate the epistemic ME SDK and show people what we're doing here? And in this case I took what I call more of a top down value centric approach to. If I am a developer and I'm building an AI chatbot, how do I, how can I use the SDK? So I'm trying to make it the most useful that it can and meet the existing best practices of using conversational eval evaluation tools to train a coach. And part of that was to get the data set configured and I'm using the same dataset configuration as deepaval, which is an open source evaluation library that I found really useful. And so in we're using deepaval and then writing our own evaluation metrics on top of it that are for personalization. And so the first thing you can see here is that in the dataset configuration I can filter it by user cohort. We have some examples here, like fitness enthusiasts. And it'll basically save, it'll store conversations. And then there's the router intent, which is when the AI coach is interacting with a given user, it interprets an intent that the user has. So maybe they want to do some research or analysis or something like that. And you can filter for conversations that look like that. And yeah, then you can just kind of browse through some different example conversations. These are like really simple, these are all generated, but I'll be importing more substantial conversations. And so the structure of the conversation is like the back and forth between what the user is saying and the coach is saying. So I have these kind of like chat bubbles here that indicate that. And these are just, let's just have one or two turns, they're called turns within the conversation. A real dialogue would have multiple turns. And I'm really thinking about the case we talked about, Robert, with like the SOAP loop, like where user uploads data, we ask subjective questions about the user to learn about them. We populate a belief system about that user, we make a recommendation and then the user selects a plan for that recommendation. All makes sense so far.

Yeah. Could you real quick for the audience that didn't catch the boshoi pond define soap real quick, I think you started with yeah, if you could define that real quick just for the audience because I think that'll be key for them understanding our design pattern here.

Yeah, I guess a term used like in a doctor might use in an evaluation of a patient. So subjective objective assessment and plan and yeah, and it just was some language for what how we were already thinking about the cycle of a user uploading some data. It could be a PDF with information about them, some it could be the health data or whatever. That's the objective profile. Subjectively, we're going to inquire about various subjective qualities of what's going on based on the. This kind of thing underlying beliefs. And then the assessment is the. That's the recommendation system from like machine learning point of view. We assess, we recommend and then we, we choose a plan with the user based on what they actually want to take from the recommendation.

Cool. So thanks so much for explaining that. And real quick to level set for the audience. This is a core feature of most developer facing tools and certainly epistemic Me is a developer facing tool with the outcome of hyper personalization and right now in the niche of health and longevity focused AI coaches. So when a developer comes to our website, ideally what they see is this SDK playground with some more design sprinkle in here where they could really see, oh, this is for them. Oh, this SDK helps them get the best dialectics for the best hyper personalized experiences for the health and longevity use cases for their end user. And also getting familiar with how the escape playground will help them be a training grounds and simulation grounds to get the best dialectics or turns as Jonathan was saying. And then on the right you can see the data model or the schema. So getting a developer very familiar. Like they go into our page and I'm imagining they see, oh, this is for them, I got to build an AI health coach or AI longevity coach that's hyper personalized. How exactly do I do that? How do I make sure it doesn't go off the rails and provides relevant recommendations that a user will actually adhere to that will be useful. So then they see this, they're like, oh, this is really cool, this is for me. And then they get very familiar with our data model and hopefully they go to our GitHub and pull it down and start playing with it after they play with it through the SDK playground as first so sort of value step in their journey. So with that context, additional question here, could you explain the data model on the right?

Yeah, the context on the right? Yeah, yeah, yeah. So right as, so as the conversation is happening, this is what the user might see in the chat browser. But beneath that the chatbot's gonna have access to context about the conversation and about the user. And so we're just scaffolding information is going to be there. I think of particular importance is this placeholder for the belief system. And say in this interaction here, the user asks a question, the coach gives a response, the coach may respond with a question at that point to learn about the user to actually like fill in beliefs or it may extract it from like whatever the user is asking may have some previous context from before. And that's one of the main goals here is just like on a turn by turn basis that we can look at the beliefs that are extracted and we can see towards our learning objectives. The coach is actually the one that's going to have learning objectives and the coach is going to say I want to learn about this user's habits or like sleep, diet and exercise for instance. And that's the coaching learning objective and then it's going to ask questions about those habits and it's going to extract beliefs about that involves those habits.

Got it.

And eventful. Yeah, so I don't think I pressed this big red button yet. And this is actually like the whole point of this page is that it's going to run an evaluation. And this uses Deepaval's evaluation toolchain to talk to OpenAI and then it evaluates whether the conversation achieves its goals. In this case, I've created three custom evaluations. One for user identification, one for belief evidencing, and another for belief verification. And then it runs through the evaluation loop and it gives a score about how well this is doing. And so the whole point about this interface here is that you can test new prompts for your coach and see if they return better responses and that the conversation improves so that the evaluation metrics improve. So it's an improvement loop to train the AI coach.

Very cool. So the goal is developer tasked with building hyper personalized AI coach goes to our website, sees this SK playground. They click run evaluation based on some pre existing data we populate that speaks to their use case like you have here. And then they could see, oh, this is the thing that will help my coach be better and better and learn more about users and get into that loop of perfecting these evals. Is that right? Am I hearing that right?

Yeah. In this visualization here, you just want to see these numbers go up and the bigger surface area on the radar graph be visualized. The better that is, the better your coach is according to your evaluation metrics.

Cool.

And these are specific, like you can. You have a lot of different training metrics for a conversational chatbot or a coach. And these are specific to some of the personalization metrics we've talked about using epistemic me for. Yeah, you just have somebody come, let's say you don't have a self identified fitness enthusiast yet, but you're just talking with somebody. Can you start to identify from their demographics or from the belief structure what user cohort they belong to and start serving them content and questions that are that meet that identification. So that's one of the, I think easy to conceptualize goals that you don't need to know much about philosophy to get the idea. And then just starting to disambiguate what user cohort somebody's in. Now we're like looking at some of the different belief extractions and so we're extracting beliefs in context here. And then we're going to the trying to learn about the user. So belief evidencing means, like, how well do we understand how the user evidences their own beliefs? Does that make sense?

Yes.

And I guess I'll give you an example. Like in health, let's say you're going to. We're planning to recommend a plan that recommends a certain. Like, we were just talking about four by fours and improving VO2 max. Yeah, let's say we're going to want to back into like, how did this user come to the belief that doing four by fours will increase their VO2 max? Is it some socially geared evidence? Is it something that they have read papers about? And how do they ground their own beliefs? Maybe they're an athlete and they just know. Yeah. When I do four by fours, like my VO2 max goes up. It's like very experiential. Right. So understanding how people think about these beliefs and how they evidence them is one of the goals of understanding users.

Yes. And that leads into being able to recommend the next best thing or ask the next best question to learn more, or give them the best habit change or recommendation for exercise, diet, whatever that use case might be or be easier. Did I get that right?

Yeah, exactly. Because the more you understand how they ground their beliefs, you can ask better questions, just like you just said. And then the recommendations that you can give are going to meet how they think about these beliefs and more with more fidelity.

Yeah. Just for the audience. One of the ways that I ground our use case is think about a real fitness coach, like a human one that you would call up and ask them to coach you. Why would you go to them? It's because in between knowing that you should be healthier, sleep better, better diet, exercise more, in between that and actually doing it, the knowledge is like a lot of uncertainty. So you go to a professional, or in the doctor use case, you go to a practitioner with a license, and then you learn from them and they learn about you to be able to understand your underlying beliefs around health. Good use case was my brother used to think before I showed him some podcasts on sleep. He used to think that if he ate a lot of food right before bed, he would sleep better because he would have the food coma effect. And so when I showed him scientific research that backed that, hey, when you eat right before bed, your body or your resting heart rate goes up because your heart has to work harder to convert that into useful energy. Now you're in a fed state. And so for my Brother, he's a PhD in biophysics. Super intellectual guy. But he didn't quite have that data point, especially not from the academic side. So when I presented that to him and I understood the root of his beliefs, the evidence was his experiential experience of hey I food coma at night if I eat right before bed. But he wasn't tracking restorative sleep. He didn't really know about that and he wasn't tracking data over time essentially like his personal data. And then now I presented a new evidence point and fast forward two months later. His sleep is better than ever in his whole life. So it literally transformed his life just getting to the root of that belief. So basically think about a fitness coach or longevity coach or influencer that has millions of followers. They cannot feasibly from a human point of view, one on one coach every single person with their philosophy on longevity and their philosophy on health and how they might recommend diet, sleep, exercise, because they just don't. There's not enough infinite time out there. Right. That's just not going to happen. So that's why the AI health coach that gets to the underlying belief of the user maximizes learning rates around that to get to the evidence behind their beliefs. And then we evolve this belief system for the user and then we can also see that by cohort health conscious professional or maybe pregnant women or whoever the cohort is so that the AI coach is hyper personalizing experiences. Okay. I just wanted to give that kind of more human example to give color context. Jonathan and saw you.

That's great.

Yeah, sorry, I saw you were pulling up some other things too. Did you want to go over those?

Yeah, I wanted to show. Because the idea here is to meet the best practices for doing. For making a chatbot the chatbot better and then using the evaluate the personalized evaluations on top of the existing ones. And I wanted to show you this was a really good article that was released just last week sometime about feel guide to rapidly improving AI products. Send it over to the don't Die team. That was really solid. And yeah they. So they recommend building your own like every company that builds their chatbot should have its own sort of analysis tool that's like home built. You just pull up each of the conversations that are happening and then you analyze them. But he really put an emphasis on not using generic metrics but letting domain experts review the conversations with you and then figure out like what the key errors are. And I think these are called critiques. So if you, if we're looking at like a conversation that an AI coach is having with a health consumer we would be looking at those and we imagine there's a, maybe a doctor looking at it, maybe there's a health coach looking at it. Any kind of expert that can really add value into what's wrong, the response that the coach is giving and then we work to fix those. Right. And so these are some examples of error categorization in their use case here.

Very cool.

Yeah. And I thought and then what they go on to show how then you can also make it easy for these experts to start, edit and curate. And so like Deep Eval's confident AI also has tools like this. But essentially you're like you're building these interfaces, you're creating a data set of what chats are good, what chats are bad, what's wrong with them. And I wanted to bring these ideas into something that's easy to use for the SDK playground too. This, it's the basic idea what's going here. But I just, I want to talk to you about some of the features I was thinking about added, adding and see aligned on all that.

Yeah, let's talk about it.

Yeah.

So that Google Data Viewer looks like key feature. Right. To bring in that domain expert. Right.

And I think here you have the same kind of data. It's laid out a little bit differently. It can always change this and there's just like a set of conversations that you're wanting to sift through. And eventually these are going to first these are going to be synthetic conversations. But then when we start having conversations with real users and we pull them in here, that's where this is where we want to view it. And I think the word turn is like there's multiple turns per conversation is what the word means. And so I thought on each turn we should have a way to create a critique and that we might consider even using epistemic me to create like Personas of an expert doctor. And so the expert doctor says this response from the coach is not accurate. And that's my critique. You remember that question, the paper that we read from I think Stanford and Google that had that approach of predicting user answers to questions.

Yeah. So some context for the audience. There was a paper that came out, if I remember correctly, that was a collaboration between Stanford and Google to they basically were trying to predict basically what the users would say next. And they had a series of role based prompt, fine tuned prompts where, where AI played a role of a psychiatrist versus a psychologist versus a behavioral specialist and many other roles around answering what does it mean to be human? Essentially, I'll call it human expertise. And they were able to get. How much was it? I know it was greater than 80, but I don't remember that number off the top of my head.

And then 80%, 85% accuracy for predicting answers. Yeah. From a trained data set.

Yeah. So basically you're saying, Jonathan, perhaps we could build a feature within epistemic me that's of that flavor to help the developer and the domain expert more easily run these evals, Is that right?

I think just think about the developer and think about those experts being there as like an, like a LLM version of a judge that can fill in for their experts before they're ready to hand it off to the experts. That's, that's one idea I was thinking about.

Oh yeah, okay. You know what's really meta is if we built self models of the experts themselves with their own knowledge bases.

Yeah, that's what I was saying. Like we just, you have epidemic need to create stuff, some structures.

Oh, that is what you're saying. Oh yeah, I think that'd be super cool. Yeah, for sure. Like instead of a slightly generic fine tuned prompt engineer method, like why not make the expert based on their belief systems and their own knowledge base? Like that makes a lot of sense. Okay, that's cool.

Yeah. So it could be a holistic tool that way because the expert perspective is giving Mackenz critiques and they can look at the whole conversation and critique like that. The plan that the user ended up with here is they're not likely to stick with. For instance, that might be like a therapist level critique of a conversation and we could have these different Personas that offer you critiques and offer you ways to like make your evaluation score higher based on how it needs to be improved. So that's one idea. But I think what I really want to see here is that throughout the conversation we can go turn by turn and the user can see like the context and the belief system evolving as the conversation is happening. So like at the beginning of conversation we maybe know no beliefs about a new user and then as we learn more about them, the beliefs populate and we have written deeper context about the user. So that's one of the key things I'm trying to show in the SDK.

Yeah, yeah, I'm now thinking about it with that lens of the data viewer that we're defining as that sort of visualization where we could do the error analysis and bring in an expert very easily. Yeah, I think that makes sense. I wonder if it would be useful to developer if we had here is not just the next best question to learn more about the user or the next recommendation, but what if we had here's the top three and here are the scores and why that might be interesting as well. So then they could see in that loop like what were the other options and perhaps pick a different one. Allow the expert to easily view a screen of oh, these were the other options. I like this one better even though that one was the one that shipped or was leveraged.

Yeah, I love that. Yeah because like that rank choice and I think we could return for the SDK like here's some questions that you could ask and they're ranked and you could on a certain sort of per dialectic turn basis if you could give feedback on this question's better than this question and so on.

Yeah, I'm really now thinking about it and we'll add this article to the show notes. It's a really good article. Here's a quote from it. Error analysis is the single most valuable activity in AI development and consistently the highest ROI activity. And the whole thinking is if you could instead of having an engineer talk to a doctor and be like, hey, how should we fine tune this prompt? Or what do you think of these terms? Slash dialectics, whatever you want to call them. How do you think that went? Just give the doctor or domain expert the tools to annotate these turns themselves. And so now I'm thinking about that user's experience, the doctor's experience, and how do we make it as easy for them as possible to provide their expertise? That's why I thought of the hey, what if we just gave them the ranked list of questions we could have asked but didn't and then give some quantitative data as to why. So just opening up that conversation for a sec, maybe we could brainstorm how might we make their lives as easy as possible to help fine tune the AI?

So there's one thing that maybe the you build the developer annotation as one tool, but then there's other tools that let you curate the data set. So just saying that once the data set's in good enough shape, you can just put it onto confident AI or a number of other dashboards that like are already built for expert curation. So part of our job there is just to make it easy to create a good data set.

We're just taking turns. How would you rate that turn?

Right?

Corrections Global I need a forward I'm reading domain expert to tell me or a Podcast host, domain expert to tell me don't switch being on mute while you're talking. Don't be on mute while you're talking. That this is how we would rate that. Cool. And any other thoughts to wrap up the SDK playground conversation or anything else you wanted to talk about there or line up?

Yeah, I think it's in a good place where you can see kind of like where it's at. But then I'd like to imagine the visual interface that is just like quickest time to value for our developer. And so I'm thinking about just like how to display the conversation and then annotate the conversation as a whole. Especially the conversation as a whole. Just annotate it with here's what's wrong, here's how to make it better kind of thing where you can like auto generate some prompts for your coach that would improve it. But then I'm also thinking about especially the visualization of how beliefs are learned during, over the course of an evaluation. And I was thinking a git diff type of basis like where here's the belief system and when we started, here's the ones that were added, here's the ones that were removed and you can just see what we learned about the user in a given session, the goal there. So those were the things I was thinking to focus on next.

Yeah, honestly, if we just get a visualization with some dummy synthetic data, get it in a prototype like over to our ICP devs on the Don't Die side, we could probably rapidly get feedback without even hooking it up to anything real. Like just the what is the thing that would get you to grok what we're doing as fast as possible and eliminate time to value. I think that would be a good path forward. Perhaps in our Wednesday call we could brainstorm the VIZ a little bit more. Cool.

Yeah. And I think the whole goal, I think with Don't Die is just to get them a baseline data set of synthetic data, that the thing works pretty well and it can demonstrate that we can go through one of those SOAP evaluations and the different context of their app. So the semantic router that I was showing on the left hand side, you have one of those for biomarkers. You'd have one of those for their Don't Die score. If a user goes in, they sync their data for the day, it creates their Don't Die score that just kicks off a session. Right. And the user can choose to answer like habit, subjective questions or whatever, you know, through a cycle Every time the data is uploaded. And so the semantic router on the AI coach, I think we can. Okay, I think we can tangibly deliver value by having here's a synthetic data set. Here's a way to train your prompts and make them better and just plug them in and then redeploy the app. That's where I'm trying to niche their development experience.

Oh yeah, we should get that up and running as fast as possible and get Greg in a meeting and show it. I think this is the best way to move forward on that. So I'll action on that after the call. We'll coordinate on. Let's just have a meeting on the calendar. Come with a prototype because in the last em don't die Team Sync, we talked about that as follow up. So we'll take that. Cool. Anything else on SDK Playground before we move into marketing meets development on the website?

Yeah, I think that's it for now. I think a design session on Wednesday would be great.

Yeah, I bet we could easily take. You had a really nice the dialectic designer. I felt like at a high level the lo fi mocks were like that's pretty good. Let's just hi fi it and then meet our website design and that'll probably be pretty good.

You're just looking at Cursor's interpretation of my ExcelliDraw.

Yeah, I think, yeah, it was good for showcasing the functionality and what kind of data you'd get. So I think we'll put a rev on it. So first off, I revamped our logo and made it more neon to match everything.

Oh, new I wasn't there yesterday. That's cool.

Yeah. And so it's now a fav icon. So now it's in the browser tab too. So stands out. Although now looking at this, it probably needs to be more like there needs to be more pop with this. I'll do something with that later. I'm still going back and forth. I've just left this as a join the wait list now to try to induce fomo and then I have a join our community secondary call to action. The tagline is on. The hero section is understand user beliefs, hyper personalized experiences. And I moved our original like very head grabber sort of tagline down here. If we don't understand ourselves, how can AI understand us? Because we're really trying to appeal to that ICP that we've been discussing this whole call. But also we want to appeal to a secondary ICP or perhaps even the first one who is Mission or purpose driven. In this age of superintelligence where everybody knows we have to solve AI alignment, anybody working in AI definitely knows we have to solve it. We want to track those people and tell them we're looking to solve it. And also here I was thinking about we really want to show them how we're trying to solve it. So okay. Beliefs lead to hyper personalization. It's the toolkit for quantifying subjectivity and belief systems for true personalization to meet your users where they need you in real time. And then just this little FOMO area and then this understated subtext to just talk through. Oh, AI alignment, SEO personalization, SEO. Cool little GitHub thing added domeda logo. Oh actually yeah, maybe I'll stay here. Let's go kind of section by section. Add. Let me get get your feedback and I'll take some notes on the side here.

Yeah and so I think one interesting thing about this is it's different. It's a for a different audience than I think our last website was which was like we were more for people trying to contribute to the open source project the last one. But this one seems more geared at developer icps. And yeah my first thought reading this is it just it should say for conversational chatbots somewhere because like how else do we know who the audience is and what they're trying to add hyper personalized experiences to? Makes sense.

Yeah. Okay. Conversational chatbots. I'll make sure that's here. Maybe something like to really bring your conversational chatbots to life. Something like that here. I'll think through the copy. That makes sense. That's good feedback. Any feedback on like this. The FOMO inducing elements of is now in beta. Join the waitlist. Thoughts on that? I'm really thinking about that. So when you click this you go to a waitlist form and this way I just get people into our CRM and right now because the website is targeted at that ICP and we're not at a point where we have product LED growth and sassifying things like out be in a few months time. Fingers crossed once we get this rep in with AI Brian and hopefully some fast follows. That's why I'm thinking like putting people on a wait list is probably the best for now so they could signal intent. And then I have this secondary join our community tab where we're really also trying to heavily curate our open source community. So. So I have a few additional things here because we went back and forth on should we just toss the open Source slack into here or should we have an intake? And I think where I landed over the weekend when I shipped this was let's have an intake because we want people that are there to really want to be there. And this early in the community I think that's important. You don't want the wrong sort of icps. You want like heavily vetted. These are people that want to build the future with us and share our values. So on this like I have some questions towards that they could tell us what they're interested in, their intent, some like what drew you to this? What would make this community meaningful to you and that way we could vet intent coming in. Thoughts on any of that? Cool.

Yeah, I do the community like intake funnel. I like that But I, I think I feel like a little bit of an identity crisis here about who is the audience that we're talking to like and on one hand it's jumping to this isn't this is an open source project. Use it in your personalized chatbot today. I think if we're trying to. If we're trying to have a conversion activity for the developer who wants to try an open source project when it's ready for them, then that's the form below, right? Yeah. And then the other person looking at this is you're a developer and you think this is interesting and will it help us apply philosophy to AI and create hyper personalized experiences. Those two different people, two different audiences. But I'm. I'm not sure that comes across if you were to user testing on it.

Yeah, I'll test it with our current open source contributed and see what they think. I almost. I had another thought and we can go back to the drawing board on this. I had another thought of maybe we have two landing pages one for each. One for each icp. That that's very clear. It's for them. That's. Yeah, that's been hard for me to parse out because then on the go to market side for this with hey you're the AI health coach developer developer tasked with building this hyper personalized conversational chatbot for the health longevity use case. Perhaps we just have that as a different entity or something outside of epistemic me. I'm not quite sure. Yeah, I'd love some riffing. Any thoughts on that? Like thoughts on two landing pages and if so we want to rip it out as a different entity then to package this product.

Yeah, I guess you haven't joined our community button. There's not a separate page for team and community or the team page meant to be that because I don't know. Yeah. You can have a million landing pages on a given website asset.

Yeah.

And here's the link that we send developers and here's the link that we send potential customers. And those are two different things. Yeah.

Have you seen any good examples of that out in the wild?

I think they're usually. Yeah, like different.

Different. Okay.

Yeah. I guess like GitHub is usually the call to action for an open source project. If you want to start contributing. Then there's contributor guidelines there.

Yeah.

And so I don't think there's anything wrong with having join our community to funnel people to Slack. But yeah, the real call to action for the developers go to GitHub and check out the code and check out the contribution guidelines and start downloading and forking, stuff like that.

Yeah. I have it understated in here. I think the GitHub icon at this point is quite like people would see that and they would know, oh, I don't care.

No, I like that. Yeah.

I saw our stars went up by one. It might be a friend that I sent it to.

It sparked him in the meeting I had today that I was telling you about.

Yeah, it could be that too. It's fun to just see this number go up though.

Yeah.

Now we gotta. We have a very obvious target. That's cool.

Okay.

So I took some feedback on that. I'll do some research. I think that's my takeaway on how other open source products projects with a product that they're trying to go to market with. How do they handle that? Is it two separate landing. They just. I'll look for some inspiration. All right, moving on. I think this part probably doesn't need much.

Yeah. I think it'll be additive and we can add a case study there so that it can see if it doesn't say that we're using it with an AI coach, which is the most key thing to get across.

Yeah. I would like to actually show a screen even like a gif of AI Brian in action here. That's what I'm imagining.

Yeah.

Once we actually have more users and reps through with it, I think that'd be really great. Cool. So that's. Yeah, yeah, we're gonna make it happen. Starting with the Greg meeting this Friday. We'll figure it out. So I still love this question. Yeah, I think good. The signals like our intent actually looking through this, I should probably make this our accent color just so it's different. What do you think of this section.

Yeah, so that's what I was thinking of the top section. The key thing to get across here is like, what do you mean by personalization? What's. What's the thing I gotta have in my chatbot, my conversational chatbot that I don't have now? And then I think this section starts to address it.

Yeah. What I was hoping they get is that's why I started with this first. Is not just like understanding your users better, it's understanding their beliefs. And like, I want that. And I say belief, you hear in the secondary line. I want that to see the developer's thought of, oh, this is a different angle on personalization. So I'm hoping they see beliefs and then they scroll down here and then it's already in their heads that, oh, that's the difference maker with this and anything else? Any thoughts on the copy?

Let's see. So what are we comparing to current AI approaches versus AI plus epistemic me? Right.

And yeah.

So yeah, this is where I think some like really strong positioning is. Would suit us. Well, just to make it like clear, this is the old narrative, this is the new narrative kind of takeaway.

What would be the strongest is if we had. So actually what I was trying to. Maybe we even. I don't know what I was trying to do. Skipping ahead here, I was trying to do this with this section and this. Eventually we will just replace this with the SDK playground. But because that's not ready yet, I wanted the user to get an idea of what does actually look like when it is with epistemic me versus without. And I generated this with our intern svelte and it from this copy paste synthetic input. Basically I wanted them to see this is the width without. This is the old way, new way. This is what I intended with this section. But I'm now thinking maybe this is too far down. Maybe it should be right here.

What do you think with and without this? I love this section.

Yeah.

And then I thought like, it's just. Yeah, let's just make these kind of like things we can demonstrate in code.

Yeah. So what do you think of having this section right here, like right after this? Because it's okay. Here's the before after. And then they're going to be like, okay, but like, how can I see this in action? And boom, it's. We see it in action.

Yeah, I think that's what I've. I'm thinking, thinking about distinct chatbot capabilities. The ability to ask your user the next question to learn the Most about them. The ability to predict how your user will answer given questions based on cohorts. The ability to place a user into a cohort with specificity. Measured specificity. The ability to understand how your user is learning through the chatbot with measurement Epistemic emotions. Like those should all be things that a developer today is like, what the. I don't do anything like that. Should I be doing that? But I think that's where we want people's heads at.

Wait, what was so ability to predict how your chatbot will learn. Oh, will respond to a user.

I'm just writing this down.

The ability to place your user into cohort with measured specificity based on belief systems. What were the other two that you named there?

Yeah, it was the next best question. Yeah, Predicting the next answer. Yeah, those are two. And then the last one was the epistemic emotions to like measure learning progress through a session.

And then ultimately all of those things we're saying are difference makers to maximize learning rate for you as a team through this chatbot, but also make sure that it hyper personalizes recommendations and maximizes value to the user with their recommendations.

Yeah, yeah. I think that hyper personalized recommendation, as in this is you're building a recommendation system. This recommendation is hyper personalized. It's higher. It's got a higher success rate of having the recommendations be accepted right by the user.

I like that how you just called it a hyper personalized recommendation engine. Because I think that will grok with them. Is that's essentially what a chatbot is doing.

Yeah.

So that makes sense. Like we. We'd basically be creating this new category of there are other.