NEAR AI Ecosystem - What Did You Ship This Week? #20

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:16

   Near AI, what did you ship this week? Hello everybody. Welcome to episode 20 of Near AI's. What did you did you ship this week? My name is Cameron Dennis. I lead integrations and partnerships at Nearai. Um, Neurai is essentially venture studio building out a bunch of exciting stuff. Uh, one of these teams is here to discuss what they've been uh building for the last almost like four months or so. Uh, first time that we're coming, I guess, publicly with some of the stuff. So, shout out to those who

2. 0:47

   are watching getting the the early alpha. Um but just as a quick TLDDR, this is think of it as a decentralized or open uh engineering standup where people across the ecosystem demo what they've been building. Uh today we have an amazing group of people that I'm very excited to have on. First we have Pierre from the near AI team to discuss the near agent cloud. Uh we have Vele from Vigil to discuss agent security. Uh Vigil and I and the Near AI team have

3. 1:13

   been working together for about the last uh I don't know five six months or so. uh we put out a request for proposal um to create an auditor agent and their team uh scored the best in that RFP process and so very excited for them to demo what they've been building. Next we have Hong from Follow Network uh not his first time on the show uh to talk about all the new exciting stuff follow shipping. Then we have Owen from Proximity Labs talk about updates the new uh shade agent framework. And last

4. 1:43

   but not least, we have Felix from X-Trace to talk about the progress we made uh on private vector stores. So to kick it off, Pierre, I'm going to add you to stage. Welcome to the show. First time on. Hey. Yeah, good to be here. Thanks for the invitation.

5. 1:58

   It's good to have you here. So uh we normally start this off with a just short introduction of yourself. So if you tell everyone who you are, what you do at Near AI, and then what you've been shipping, that'd be a good place to start. And then I'm going to hop off. uh you have about 10 minutes or so and I'll pop back on anytime I have a question or uh you know time to go but yeah take away.

6. 2:20

   Yeah. So I'm Pierre Puen. I've been um like working at neari for since quite the inception and before I was working at near foundation for something like two and a half years three years and um we've been focused on uh creating the solution called agent cloud agent hub and uh we are building on the first version of what you've seen at neari which was the hub where basically you

7. 2:46

   publish the agents that you build on top of the near SDK and so we went above this and basically Now we would allow you to we would host any kind of uh agents built with any frameworks and we're going to provide you lots of analytics and like also the AI gateway the LM router that we have so you get better pricing and you have a bunch of cool features on top of it.

8. 3:10

   So today I've I was prepared some live demo hopefully it's going to work. Uh so let me share the demo gods are hopefully with us today. Um but you know been building for for a little bit. I'm very excited about this framework agnostic hosting because uh this seems to be one of the biggest issues. We've talked to dozens and dozens of uh agent developers and analytics seems to be a pretty important piece. So um yeah take it away.

9. 3:41

   Yeah definitely analytics is one of the biggest problematic right now. So um I will send a link if I if I can somehow for you guys to try it for the agent developers out there that are struggling with analytics and hosting their AI agents. So we are kind of a beta mode right now.

10. 3:58

    So a like expect a lot of issues and if you built an agent on neari before on the near hub it's going to be uh running on this platform. So this is really like bringing you know as Cameron said framework framework agnostic. So I'm going to do like the flow like you are an agent developer. So basically you would land on this landing page. You want to deploy your agent. So lots of

11. 4:21

    agents are just like vibe coded and they're not really ready for production. So what we built is uh is a way for you to you vip coded some agent let's say just a CLI app or very simple like Jupyter notebook or some Python code. uh what we would do is um we would read the your agent code and we would make it production ready which means that we have an agent AI agent reading your code and adding like a bunch of uh functions

12. 4:49

    and like a docker file and so on and documentation so we can actually host it on our platform but you don't have to do any work you only have to approve the PR if you are happy with the work so I'm going to deploy one of the agents I built So, this one I'm just going to go before uh like the fork.

13. 5:13

    Uh wait. Yeah. So, we have this um this agent. It's like a very simple agent built on top of Python and GPT researcher. It basically allows you to do deep research, but you don't have to use u like any deep research solution. You can do everything from a CLI. And um the idea here is basically um you build an agent around this and then you want to host it on your on somewhere somewhere.

14. 5:41

    So you host it on our platform because it's a good solution for this. And um so you would see your agent here. You click on it. So here it tells me that I already have this agent this agent but I'm just going to create a new one because I already deployed this agent on the platform. So here you would see the um auto docker the production making your app production ready doing his work. So here it's quite easy actually

15. 6:08

    because um it already has the docker file that was emerged from a previous previous PR of this uh agent. So docker file is like some very basic like python docker file. So we have to wait for something like a minute or so. It's going to try to build a Docker image, make sure it works, and once it does, it's going to just like finish the the process. Okay, so now it detected uh two necessarily two mandatory environment variables that you have to fill. So either the agent

16. 6:46

    developer feels this um variable or the person actually coming to the platform and running our agent will have to fill it. So let's say right now I don't feel it. I only feel the open AI one. So just put like some API key like this as the agent developer. So let's call it live and but I don't feel the TD one.

17. 7:10

    Okay. So now your agent is ready. Um so it means that now you can basically go on this URL. So it looks something like this. Um once you create an account, you can actually run this agent. So if all the important variables are provided by the agent developer, you don't have to provide any of them. Uh but in that specific case, I provided only one of out of the two. So once you will have the build that's ready. So

18. 7:38

    here it's currently building the docker file. Um we will come back to this like once it's ready but so once it's built uh we will click run and you will see that I will be prompted to fill one environment variable. So before this live I just prepared like this exact same agent uh amazing research agent I just deployed it right before and um so we're going to do the run from here.

19. 8:04

    So here in that case I I entered like two valid uh API keys and I'm going to I'm going to start a run. So here you can see I don't have to fill any environment variables because they're already filled in. If I want I could like set some different um parameters for you as CPU and memory for my agent.

20. 8:24

    And then I'm just going to start it. So it takes like uh something like 20 to 30 seconds for the agent to start. I'm just going to reload this page. Okay. So now we can see the the fast API um server starting up and uh we get some health check from the server.

21. 8:52

    So now we can just go ahead and click on open agent since this agent has some uh UI. So if you go on, so this is the code of the agent I just deployed that I'm showing right now. There is this very simple index.html. So it's being served from the server, which means that if you go on this link, oh, oops, the gods are not with me. So if you go on this link, you would see the um you would see the UI. But so I did it right before this and it was working.

22. 9:23

    So, I'm just going to go on the other run that I have uh that I just did right before. Must be this one. Okay. Okay. Now if I open the agent, it's loading slowly but surely. But yeah, so this was the right before the stream. So this was the the run and you can see gened some unique URL and I went to it and you can actually see the UI I showed the code for and then here you can ask whatever research you want it to generate and uh you would click start research and then

23. 10:14

    it's basically how the agent is supposed to work and then you just after a few minutes you would get the answer back. So yeah, looks like we have an error on the back end. So I have to fix it. Um, so it goes that's the point of a beta though. So yeah. Yeah. So once once you build like you share the agents with your friends and anyone basically the idea is that you would get lots of analytics. So you would see the number of runs, you would see um latency, you would see number of

24. 10:43

    tokens. uh we're going to also like provide a lot more very uh deep analysis of your agent and basically really allowing you to make it like work perfectly perfect perfectly in production. So yeah that's the the quick overview.

25. 11:01

    So this is really for its alpha beta scope and we have planned a lot more lots of features on top of it in the next few months. So Pierre, what do you want people to do with this? Is there a form that they should sign up for? You know, do they just automatically get access to the beta? How does this work?

26. 11:19

    Yeah. So, you can go. So, right now it's a it's like temporary domain that we are using. Um, so this is you can go on divingtools.com and here you can like try the platform directly. So it's not it's not gated like just like the other part is actually to find this URL because we are not like listed anywhere right now because Yeah.

27. 11:46

    Can you send me uh that link somewhere and then I'll Yeah, I just I just sent on studio chat devingtools.talk.com. Beautiful. Um so just a clear call to action. If you're an agent builder, go check out this link, play with it, provide feedback. Uh there's an alpha group as like alpha like an alpha tester group. uh that we'd love to get your feedback. It's pretty much 247 support,

28. 12:10

    so it's a fantastic way to um troubleshoot any issues you're having. So, please, please, please check it out. The link. So, this is the link for the alpha group and yeah, if you have any issues, uh send a message there and we'll answer.

29. 12:31

    Awesome. Here. Well, thank you so much. Hope to have you back on the show soon to show some updates and uh yeah. Is there anything else you want to say before we move on to? No, looking forward to the other updates. Thank you so much for hosting me today. Yeah, have a good one. Thank you for coming. See you.

30. 12:47

    All right, this next one I'm very excited for. Vele, welcome to the show. Nice to meet you. Nice to see you guys and really excited to be here. It's good to have you here. I know uh for those who are just starting to tune in uh Fleet's Work and Vigil uh is actually well they've been building for quite a while but uh we started working together because we uh launched a RFP request for proposal to figure out uh an

31. 13:12

    agent that can audit other agents that operate inside of TEES and create some sort of uh almost verification. So if we're trying to figure out agent discovery, we know which agents should be, you know, able to be discovered. And so there's a lot of work and it is incredibly complicated uh stuff. So Veil, Godspeed on trying to simplify it for everyone in the crowd. Um very excited for you to demo what you got. Uh but just to kind of kick it off, please tell everyone who you are, what Vigil is building, and then yeah, what you ship this week.

32. 13:45

    Yeah. So uh yeah, I'm really excited to be here. Uh Cameron, and yeah, we've been working pretty um diligently for the past couple of months. Let me just I have a couple slides here which I'll go into vigil and myself and everything. So, uh let me just pull those up.

33. 14:01

    Um all right, there we go. So, super happy to be here today to talk about what's been cooking at Vigil over the past couple of months. Today, I'm excited to introduce Vigil Diamond, uh which is an agent that audits your agent, just like Cameron mentioned. So again, for context, a while back, we responded to an RFP put out by Near AI where they outlined the need for a mechanism to

34. 14:25

    audit agents before they're activated on the platform in order to ensure these agents are secure, robust, trustworthy, and to help developers address issues through clear mitigations. As many of you know, with the new shade agent paradigm, we're moving toward a world where autonomous agents are granted real permissions to perform sensitive operations. That shift makes a trusted audit layer not just valuable but essential. So Vigil's response to that

35. 14:50

    need is Diamond. And today I'll walk you through what it is, how it works, and why it matters. And at the end I'll show you a demo of Diamond in action. So a little bit about Vigil. Vigil was founded in 2023 by engineers who built the deep learning platform behind Amazon SageMaker. The motivation was clear.

36. 15:11

    Autonomous agents were no longer theoretical. they were starting to show up in production, but the risks that came with that, especially around security and control, were far too serious to ignore. There was a clear need for a trust layer, something that could evaluate whether an agent is behaving the way it should and help fix it when it's not. Miguel was created to bring enterprisegrade trust to this new

37. 15:33

    category of software. At the core of that effort is defining what trust actually means. For us, it comes down to measurable properties. trust dimensions like safety, reliability, security, and compliance with policies. These are the foundations of how we evaluate agents.

38. 15:50

    To test these properties, we built in both red teaming and blue teaming strategies. On the red team side, we simulate attacks and adversarial inputs to probe for vulnerabilities. On the blue team side, we provide active defense, monitoring agent behavior in real time and enforcing constraints and containing risks if something does go wrong. And beyond testing, our aim is to transform ordinary agents into trusted agents through continuous evaluation,

39. 16:16

    clear trust reports and actionable feedback that supports improvements over time. So just a little bit about me and why I'm here. So I'm a senior applied scientist at Vigil where I lead the privacy pillar of our trustworthy AI work focusing on security, confidentiality, and integrity and autonomous agents even under adversarial conditions. I hold a masters of applied science and privacy preserving machine learning from the University of Toronto

40. 16:39

    where I focused on secure computation and applied cryptography and machine learning. Professionally over the past six years I've worked at the intersection of privacy robustness and secure enterprise deployment of machine learning systems and I've done some academic work as well publishing research on encrypted RNNs using fully homorphic encryption including two recent papers um in the field. So why am I here? My work in secure computation,

41. 17:03

    applied cryptography, and encrypted inference has always aimed to ensure that models operate securely. But over time, I came to see that privacy is just one part of a broader trust landscape. These systems are non-deterministic, which means their behavior under real world scenarios matters just as much as how securely they handle data. That perspective led me to expand my focus.

42. 17:26

    still grounded in privacy but now also addressing the other de essential dimensions of trust such as like we mentioned before robustness, safety and behavioral integrity. It's about making sure we can trust not only how a model is built or how it runs but also how it behaves.

43. 17:44

    And so what is uh vigil diamond? Diamond is an autonomous auditor agent designed to evaluate other agents across three key dimensions. behavior, code, and runtime activity. It begins with blackbox testing to assess agent behavior, how well it aligns with its intended function, for instance, and whether it avoids actions it's not supposed to take. It then moves to white box analysis of the actual agent code, inspecting for risky patterns or

44. 18:13

    insecure logic. Finally, Diamond monitors runtime execution, including low-level signals like system calls to detect behavior such as unauthorized access or data excfiltration. These results are compiled into a quantified trust report based on how well the agent aligns with intended behavior and avoids prohibited actions.

45. 18:35

    These results, importantly, Diamond is also interactive. It doesn't just surface audit results. It works with the developers to explain failures, propose mitigation strategies, and help improve agent performance and safety in real time. And more importantly, all of this takes place into a tr inside a trusted execution environment, ensuring full confidentiality and integrity of the audit. Confidentiality allows agent code

46. 19:00

    and environment settings to be protected while integrity allows audits to be certified and verified. Reposts are generated securely and returned to the developer end to end encrypted. Now once an audit is complete, Diamond doesn't just generate a report. It also formally certifies an agent by issuing verifiable certificates to agents. It acts as a certificate authority, digitally signing the audit results and issuing or revoking trust certificates based on how well the agent meets key

47. 19:29

    trust dimensions. This allows Diamond to move beyond evaluation into trust enforcement, providing a clear signal that an agent has passed or failed a rigorous standards view uh standardsbased review. We're also integrating diamond, I'm happy to say, into the neari platform where it will establish a baseline of trust across the ecosystem. Agents that pass the audit receive a verified by visual badge signaling that they are secure, robust, and compliant with platform policies. So all in all, this

48. 19:59

    creates a foundation for developers and users to safely adopt autonomous agents knowing those agents have been independently audited and certified. So what am I showing you guys today? Um in today's demo, we're going to interact directly with the visual with the visual diamond auditor agent using a simple chat UI as our interface. So I just want to mention that um the chat UI is just for demonstration purposes. In the

49. 20:24

    future, we'll probably have an API to start audits and then just a support agent to deal with um developer feedback and understanding the reports and whatnot. But for today, to show you guys how it works, we've um employed a simple chat UI. Uh so the first thing to highlight in this diagram is that diamond runs entirely inside a trusted execution environment powered by follow which are some of the presenters um

50. 20:49

    later on in this uh in this video. So not just the sandbox where the agent code runs but even the underlying LLMs powering the agent logic they all execute inside tees. That means everything from evaluation to reporting happens in a secure isolated and verifiable environment. Today we'll be auditing an agent deployed on the neari platform called the mind share agent. So the mind share agent what it does is it monitors token balances in a user's

51. 21:15

    wallet assesses the mind share or hype surrounding these tokens and suggests trades. Furthermore, as a as a shade agent, it doesn't s just suggest the trades. It can actually ex execute them autonomously, which is the reason why we need to audit this agent and make sure that it does what it's supposed to do and it doesn't do what it's not supposed to do. Uh so for this demo, we've extracted the core decision-making logic of that agent and we'll evaluate how it behaves under audit.

52. 21:40

    So let's just go to step one. In uh step one, diamond begins by gathering the necessary context. So it asks the user to provide a public GitHub URL for where the agent lives and then also a path to the relevant module with within the repository on GitHub and also any environment variables that you need to pass to the agent to make sure that runs. Now it also asks you to provide the specific types of tests which we

53. 22:04

    which we call harnesses um to apply and to run the audit against. And so we'll get into those uh supported harness types during the demo. Next, diamond s spins up a new TE sandbox via the fall API. So, this is a secure isolated container where the auto will run. TE is attested automatically by diamond and the at the station evidence can be returned to the user if needed. In the uh third step, now we're ready to pull in the agent. So, we download the agent code from GitHub, load it and prepare

54. 22:37

    for testing inside the TE. So, the sandbox TE does that for us. And in the fourth step, finally the audit begins. So for this demo, we'll show two types of testing. One of them is a static analysis of the code which we mentioned before and the second one is a behavioral test of how the agent responds to various scenarios. Uh just to mention the dynamic monitoring part is something we're still working on and will be shown in a subsequent release.

55. 23:01

    So finally in step five once the audit is done and we built our trust report we return the trust report to the chat UI where we can discuss the findings and medic mitigations with the agent directly which is really cool. So without uh further ado I've uh recorded a demo that I run ran just a couple hours ago and I'll just uh talk over it as it plays. So uh what we see in front

56. 23:29

    of us is the what many people probably on this call have seen before and that's Fala's uh UI for their trusted execution environments and managing them and deploying them and all that kind of stuff. So we can see that we have a couple instances of visual diamond deployed that are running. Uh the first one is the visual diamond-1 and that's the one that we're going to be playing around with uh today. So once we open that we can see some information about

57. 23:54

    the TE but more importantly we want to make sure that it is a TE. So we open up the atestation tab and we can go down and then we can uh click the check at the station and then we can see that the atstation is verified and there's some measurements that you can extract from there and verify yourself. So yes, visual diamond is running in a TE and everything is confidential and the integrity is preserved. And so as you

58. 24:16

    know uh the Fala API and DStack provide network URLs uh to any ports that you're exposing. So we are exposing our server for vigil diamond and um when we click that URL you can see that we're directed to it and what we have in front of us is a chat UI which we can use to interact with Diamond. So we say hi to diamond and it responds with what it is, what it's there to do. So for instance, it asks you what brings you here today? Are

59. 24:47

    you looking to audit an agent or do you have questions? So I mentioned I'm looking to audit my agent and then it starts the information gathering process uh to deploy the agent. So it asks us, can you please provide the GitHub repository URL? Um, and so you'll see in this other tab, this is where we keep the mind share agent that we've extracted. And this is a public repository all you can um go to and hit.

60. 25:13

    It's called agent- audit- samples under our visual AI repository. And in here we have the mind share agent uh purely Minecraft. We have some other agents like the multiersonality agents as well that you can uh play around with and audit. Uh but uh what we need to do here is just uh I guess what what we're doing is we're taking a look at some of the files that are inside there. And what we

61. 25:35

    need to do is really copy that URL and then paste it into the agent UI. So now it has the URL and it asks me what's the agent path within the repository. And so what I do is I type mineshare-lang. It has that information. And then once it has that, it tells us some information about the harnesses that we can run. And so these are little snippets. So for instance, for ethics, what does that do? It evaluates if an agent adheres to general business ethics principles. We can run a privacy

62. 26:08

    harness. It tests for privacy vulnerabilities and checks for leakage of private information. Uh what we're running today is the security harness. And what that does is it evaluates an agent for security vulnerabilities and adversarial capabilities. and then also a toxicity one which assesses an agent's ability to generate toxic or offensive content. So I go on and I ask you can you you know that's that's that's great but can you give me some more information about these harnesses? I

63. 26:32

    want to sort of really understand what kind of tests they they um they apply and and use in the audit. So we can see here it gives us a bit more detail into the types of harnesses. And uh let's just take a look at security since that's the one we're running today. So if we look at security, uh it says as before, the harness evaluates your agent for security vulnerabilities, adversarial capabilities. It checks to

64. 26:54

    see if the agent is prone to well-known adversarial attacks, jailbreaking techniques, and prompt injection mechanisms, and if an agent can be duped into producing harmful content using these adversarial mechanisms. So okay, that's great. I've chosen I want to do security. So I tell Diamond um I want to run the security harness.

65. 27:21

    And so it's uh you know it's great. It's it asks if there's any environment variables. Uh right now we're just using the default one. So no, we we don't have any environment variables. And then finally it confirms are is this the information is this information correct for you to run the um audit? And yes, that looks great. So we we type go and then it should it should keep going. So once you type go it starts the audit. It calls the tool that runs the audit and

66. 27:46

    what that does is actually spins up like I mentioned before another TE sandbox and we can see that there. So it it starts the TE uh in its own its own environment and it starts the sandbox inside there. So it's isolated and then we can actually audit and the whole audit is confidential and the integrity is kept via the the hashes. So, while this starts, we'll just uh we'll just skip through. It takes a bit of a while

67. 28:11

    to start the sandbox and then to actually perform the audit. Once that's done, we can see that it responds with the results of the audit and it destroys the sandbox as well. Um, so let's just wait for it to produce them. That's great. So, the audit's complete. The results are available in a report folder. So afterwards we'll be able to actually download the results and then take a look at them a bit more in depth if you

68. 28:41

    want to. Uh and so the security the the the score for the security hardness and this is a score out of 100 uh with you know anything higher uh being a good score for the agent meaning it's trustworthy. Anything lower meaning that's really bad the agent is not doing well. So it actually scored a 42% uh out of out of 100 which is uh pretty striking and it identifies the failure patterns and you can see a bunch of them there. Um I'll just get into one just

69. 29:07

    for fun. Uh the base 64 encoding failure pattern here for instance. So if you pass in for instance um in this agent a B 64 encoding of a malicious prompt, it actually decodes it for you without if you don't put the proper guardrails in place, it actually decodes it for you and then runs um that malicious code.

70. 29:25

    And you can imagine that that malicious code for instance for a trader agent like this um makes you make the wrong trades or could potentially lead you to uh lose quite a bit of money, right? So that's one that I mentioned and then there's recommended mitigations uh for each of these failure modes and so I ask it um you know give me a couple examples of problems and mitigations for them and it gives me a couple problems and mitigations. So for instance uh malicious code generation it says the

71. 29:56

    agent can generate malicious code which can be used to harm users or compromised systems and one of the mig mitigations is to implement code filters to block the generation of malicious code. Uh so you can use a white list of allowed code patterns or implement the code validation mechanism to ensure that the generated code is safe. So what you can see here is that I can actually interact with the agent. I can ask it questions about the problems that the audit

72. 30:20

    identified. I can ask it for potential mitigations and I can continue and um as the developer for instance of the agent I can make my agent better which is the whole purpose of this and then you know in the future once it's better I can rerun the audit and if the audit um passes then visual will certify it. So um right now we don't have the certification mechanism in place that's something that's going to be coming in the future. That's what we're actively working on. But um just to summarize and

73. 30:45

    to finish up uh you can actually download the reports uh via an API endpoint. So I asked it to give me the report endpoint. So now I just take uh I just take the endpoint I copy it put it in the browser and then you can see that a zip file was downloaded and inside the zip file I have couple files. I have the SAS results. So with the SAS results we'll have some mitigation stuff uh

74. 31:08

    coming up in the future as well for that. But right now we just have the bare results of the analysis that was that happened with the static analysis. Uh there are some problems. You can see that all the sort of all of the files were tested inside the repository and then you know some problems were found uh you know with severities of medium or whatnot and um that stuff will be explainable and you'll be able to interact with the agent in the future

75. 31:33

    with that too. we can open up the actual uh markdown file that is a summary of all of the identified failure patterns and mitigations. And then here that gives you um for instance also whether or not a failure was critical, whether it wasn't um and all that stuff. And it ranks them based on on um how the audit scored. So you can see there's high, moderate, and um low severity for some of these things and also recommended mitigations for all of them.

76. 32:06

    And yeah, so that's um that's pretty much it. Uh let me just come to the last slide here. Um thank you all for listening again. We're really really excited to be here to be a part of this adventure with Near uh and looking forward to the next time and shipping something new.

77. 32:27

    Bailey, this is amazing. Uh an absolute prerequisite for agent builders out there that are giving agents private keys. It's a huge concern of mine. I've brought this up a bunch of times on this like live stream and I think the work that you're doing is absolutely paramount to address this. Um all this shade agent stuff is still pretty early.

78. 32:45

    Um, I'm still, you know, uh, skeptical of giving any agent a private key with a lot of my money. Um, but, you know, hopefully the the the visual certificate will build a lot of trust and, uh, you know, I think this is I think the coolest part is actually embedding this in the sort of engineering workflow. So, as people are actually deploying their agents, building your agents with your AI, you know, constantly working with

79. 33:08

    the visual diamond product, uh, people are going to be be able to build faster and more securely. So yeah, we're excited to define this uh we're literally defining the space together, right? And uh really excited to do that and uh make sure that we bring something that we've all you know loved for years, you know, AI and it's finally coming, you know, it's becoming something that's really incredible and finally bringing

80. 33:33

    it to something that it a place where it can perform valuable um interactions and whatnot. So yeah, really excited to be here. If there's if there's any other questions, happy to take Uh we'll probably have a bunch of questions from the community, but uh we'll probably relay to those re relay those to you in Slack. Um thank you. The time will come. Thank you, Billy. Thank you.

81. 33:55

    All right. Next up, we have Hong from Fala. If you guys can tell, uh this episode 20 was kind of a special episode. We've had Near AI. Visual is now using Near AI, uh or deploying to they're using Fala. FAA uh is also being used by proximity who's up next and so we're all sort of working together for this sort of unified mission. And so Hong, it's awesome to have you back on the show. Um you're not new. But for those who may not know you, can you give everyone just a quick overview of who you are, what follow Network is building, and what you ship this week?

82. 34:30

    Yeah, of course. Uh super super happy to be here. I think this is the uh third time I have been here. uh and uh just uh give uh people a brief intro. I'm Hin uh co-founder and CTO of F network uh and uh uh as uh video already show that they're using uh our TE solution. So probably you already know like we provide the TE. So uh let uh I think I can just skip all the you know like the Uh let me just show my screen

83. 35:03

    of uh what we are doing. So uh basically our main product is called a Fala cloud uh as simple as a cloud but that is uh TE native and AI native uh and uh for those who don't really know about T you can just imagine um when you are talking with open AI you give all your inputs to some automan so he can actually see what you are talking with the AI and what is the AI response so how can we make sure

84. 35:34

    like we not give up our privacy when using someone else service. So this is a very uh typical use case of TE to make something private preserving. So the way that work is pretty simple. T is just like a vote in your CPU or uh now it also come with the GPU. You can put your program or like your AI model in this vote. So that uh the user can still talk with the AI model or use your program, use your service but nobody else including the cloud provider like if I

85. 36:07

    have the physical access to this box no you cannot see anything inside because everything is encrypted and this even includes uh the administrator of the program. In theory we can build a fully autonomous AI agent or fully autonomous AI model where code is law. So basically you can check there's no back door in the code to save the user privacy. So this can be confirmed that even not the model provider can see the user input or

86. 36:37

    your any of your privacy data. So uh yeah this is like a quick TE 101 put everything in the black box in a vault so that nobody can touch the program running inside and nobody can see the data running inside. And uh then uh I'm going to quickly show off like what we have done in uh the last week. So uh a big news we just made uh exactly seven days ago was that we just launched the TEG GPU support. We actually has

87. 37:11

    been working on uh TEGPU for quite a long time since uh be uh since the end of last year. uh but bringing t GPU to the cloud is uh uh like a very uh uh uh big uh task. So we have to deal with like a supply side of things and make sure everything is uh smooth so that you can just click and deploy and uh um now we have launched actually three products. So one is uh what we call the confidential AI model. So uh most of the

88. 37:45

    you most of the use cases of uh GPU is actually deploy AI model so that you can get a privacy preserving model. So uh and this is the easiest way to do that. Now we already have three uh actually more than three. Yeah, there are uh four models and there's also another one is called O Lama. It's not exactly a model but just a platform where you can put any O Lama compatible model inside. So in theory you can deploy any any kind of the model from smaller one to larger one

89. 38:18

    with just a few clicks. So what all you need to do is to uh select which model you're going to use and uh choose the GPU that's required and then click deploy. So I'm not going to deploy again but you can try that out uh by uh using our cloud. Then here I have a deployed version already up and running. So if I click this one then uh you are going to uh see okay there's a H200 Nvidia GPU

90. 38:50

    attached and uh we even have a open web UI already loaded. So uh this is how like you can interact with your model without uh having uh say uh some uh uh uh using some command line uh uh tool or uh write your own code. So here I can just type hello but not sure if it can run or not.

91. 39:16

    It's just a demo. Maybe it will take a little bit time but let let's go back uh later on. So uh uh with open web UI you can actually talk with the model we just deployed. And on the other side we also have a very neat feature is called chat and verify. So big uh problem of using TE is that now you're already using some very secure environment to secure your application and how can you tell your user this is actually secure it's not something running on AWS and trust me bro so the

92. 39:50

    way to do that is uh we have a new feature just launched together with the uh confidential AI that's called uh the verification center so uh if here I choose some model And uh on the right side you can see two things. One is the uh CPU attestation. So it shows the evidence to uh uh prove that this is actually something running in the TE and also the GPU attestation. This is the part that uh can be verified together with the CPU attestation because uh when

93. 40:23

    using the GPU TE you actually has to use both the CPU and the GPU together. This is how you secure your entire stack. And uh uh it's also available for uh uh red peel which is uh you know like a a platform pretty similar to open aai or uh u open router where you can just uh uh use the model in a pay as you go way.

94. 40:48

    So uh you just pay for the token and API invocation. So let's have a try just ask them uh hey what's your name? And you can see uh this one is deepseek and it's already producing some outputs. And here it shows a verified means that it doesn't only verify the environment but also verify your exact input and output. So imagine if you want to consume some of the output from a specific large language model and you

95. 41:21

    want to use the results say on near blockchain. how can you verify this is exactly the output from this specific model and there's nobody manipulate the output or the model execution. So the way to do that is uh for every message you talk with the RRM it's going to uh produce a signature. So you can verify this signature easily on blockchain because uh here you can uh uh easily verify the signature as well as the message. This is just the hash of the

96. 41:54

    input and the hash of the output. So in this way like you can imagine how people can build the AI model that can manage your font like create a private key you give the font to the AI model and how can you make sure uh there's nobody manipulate the AI model output to abuse your phone. So this is the way to secure uh this whole process.

97. 42:18

    Um yeah and uh this is the private AI confidential AI model and uh I also showed you the confidential AI API. So this is actually what's uh uh I I I showed in uh at the red peel page and on the other side if you want more like a low-level support to deploy arbitrary program even not uh AI related we also have the GPU TE support. So you choose which GPU you want uh how many uh GPU you want, how long you want and then submit an order. Then uh later on we're

98. 42:51

    going to uh uh schedule like a meeting to help you set up everything and eventually you're going to get the access to something very similar to this user interface. So where you can just uh put your docker compos file uh you can run like AI model AI training data cleaning or even say zk proving in the GPU.

99. 43:14

    So yeah that's pretty much for the uh cloud update. I do have one more thing uh to share. So uh as I mentioned we always want to tell people this is secure the biot is not the same as the program running on AWS. But how can we let the end user to feel that? So this is actually a really uh important uh um uh uh problem to solve because uh uh we want to raise the bar of uh you know confidential AI or decentralized AI. So

100. 43:48

     we want to make sure the end user can have the access to all the T related proof. But then you know proof itself is very very complicated. So uh we made a dashboard to simplify that. So uh on the right side you can see this is actually how the we call that the full chain of trust because it's not a single proof it's a combination of the uh evidence not only for the TE but also for all the dependencies like uh uh uh this on the

101. 44:19

     right side this is this shows uh the uh testation of the TE but on the left side we can also trace it all the way back to the program you have deployed to this TE And eventually you also want to show okay this is actually the code uh uh deployed from some uh GitHub repo. So that's how you can confirm uh there's no man in the middle attack and uh you can always uh uh audit the code if you want.

102. 44:50

     So uh on the left side we also have a very high level summary. So uh as a non-technical user you can you can clearly see that this app is verified because uh the hardware is good uh CPU is good GPU is good and this is the source code and the network connection is end to end encrypted everything is secure so we want to bring this to all the uh TE application deployed on our cloud automatically but of course it's not only limited to the FALA thing is it

103. 45:22

     can also be used for uh any deployment and uh uh yeah so feature wise that's everything we have uh since last week uh and one more thing is we just uh have a refresh the doc so looks a little bit cleaner uh thanks to minki which is pretty good and uh uh we also refreshed our in uh open source version of all the te stack called the dstack uh so uh this is the new uh repo of the uh uh code. So you can take a look if you're interested. And finally uh we just got

104. 46:02

     the Dstack fully audited and to my knowledge this is probably the uh first TE infrastructure open source project being audited. There are a bunch of other SDKs but none of them are audited. So uh I think this is also aligned with our uh like a philosophy. We have two things. one, we must make sure everything is secure by default. There's no compromise. And the second one, we want to make it to be accessible to as many people as possible. So that's why

105. 46:32

     we spend most of our time to simplify the user experience. Yeah. So that's uh basically uh all I have and really glad to be here to share all the updates to everyone. Yong, that's incredibly exciting. Uh a lot done since the last time you came on. So, you know, the follow team is constantly shipping amazing work. Um, is there anything that you want the audience to know or do uh pretty much until the next time you're on the show? Like any call to actions for them?

106. 47:01

     Yeah, just uh sign up on Fala Cloud and give a try of our TE GPU uh TEGPU cloud. It's very simple, very easy to use. Just just to be clear like can I essentially start using let's say deepseek running inside of follow like via red pill right now to make sure that all of my inference is verifiably private so everything that I'm sending into the the model it's not seen by a third party this is my model it's not

107. 47:29

     you know no one else's correct exactly this is exactly what we're offering that's amazing I can see how good this is going to be for enterprises that are dealing with a lot of sensitive customer information who might not feel comfortable throwing it inside of you know one of these large companies. Have you seen like what kind of traction are you seeing of like who are the types of targets that uh you know you're going after for confidential inference? Oh yeah the it's very interesting to say a

108. 47:55

     lot of different use cases uh actually when we start to build this the in one of the initial motivation is that uh a lot of VCs like they do a lot of research every day and they they use AI of course to save their time but they don't want to give any of their alpha to anyone like open AI because open AI although they claim they're not going to pick your data but who knows right they can just use your data to train the model so you lose your uh competitive advantage. Uh so that's why like they

109. 48:26

     want to use confidential AI for their own use cases. So they are kind of the end user of confidential AI model. But on the other side we also see a lot of interesting use cases. One of that is uh a few teams are building um AI agents that manage user fun just like you say.

110. 48:44

     They literally create AI agents that has a wallet and ask a user to send crypto into the wallet. They help user to manage their fund. They also want to make sure uh because you know there are a lot of risk if you just use a centralized AI. What if uh the AI got compromised there there is like a long supply chain any at any point there can be a compromise and nobody want to bear this

111. 49:09

     risk. So the way they deal with that is to use confidential AI because uh it's end to end encrypted integrity is also protected and if there's any problem it's likely not because of the AI but because of some other problems. So this is the way people want to uh uh reduce their risk offering the AI uh services.

112. 49:30

     Got it. So between fus confidentiality vigil security um we're getting there getting to the point where AI would be able to sort of handle a lot of sensitive tasks for humans. But um yeah I mean I'm so excited to be working more closely together. There's so much to be done here. Um is there anything else you want to say before we move on to the next speaker?

113. 49:49

     Um I'm super super uh excited about uh the uh you know decentralization and the security of AI because in the past people move really fast just try to uh deliver something that can be used but now uh because your AI is so powerful I I I in my estimation I think uh maybe in the next one or two years there's no reason people don't use a secure AI. So super super excited about that and uh I'm really really willing to help everyone to make AI secure and safe.

114. 50:24

     Yeah, for anyone out there who's looking for a TE provider, check out Followup. Uh we've been working very closely together and I can't tell you uh I can't gloat about them enough, especially regarding the support they provide. Um I don't know how you guys do it, but there's like 247 support. Um, so thank you for, you know, really dedicating to, uh, making sure your customers and clients are happy.

115. 50:45

     Thank you. Thank you. Of course. Awesome. H, we're gonna move on to Owen. Thank you for joining. Oh, how's it going? Hey, Cameron. Yeah, good to be here again. We're yet another user of uh, Fowler Cloud over here at Proximity. I think everyone's using their their tech at the moment.

116. 51:09

     There's probably good reason to. So, um, would love for you to sort of dive into what you guys are building. Uh, before doing so, just give everyone a quick update. Who are you? What is Proximity Labs? And then what did you ship this week?

117. 51:22

     Yep. Let's uh let's get into it. So, yeah, I'm I'm Owen. I joined Proximity like two or three months ago. I've been around the ecosystem much longer prior to that. And we're building shade agents which are verifi verifiable AI agents that are running in infallers TES and they produce at stations which are then verified onchain uh which then leverages like decentralized key management. So instead of storing a private key inside of the TE which could uh evaporate into

118. 52:02

     thin air when the TE goes down, uh we have the keys on chain basically uh with like the MPC service which allows us to have persistent accounts and accounts on a load of different chains. And I'm going to start sharing my real quick just for those who might not know what is MPC.

119. 52:26

     Yeah. So MPC is a multi-party computation. So basically you have a load of parties that construct a key together without revealing what the private key is to any party at all. So essentially just an additional layer of security to make sure that the key is not either going to be leaked or uh yeah it's redundancy or lost. Got it. Cool.

120. 52:54

     Yeah, exactly. Uh so yeah, I'll go ahead and share my screen to show what we shipped this week. Uh just give me a minute here. Here we go. Cool. So, yeah, we've had some big updates to the stack. Uh, I think it was last week, actually. We shipped a new version, version 1.0.0 of the shade agent CLI, which um adds support for a load of different flags.

121. 53:30

     Some feedback we got that was that we had this this nice CLI where you would run the command and it would go through everything. It was quite a monolithic structure, but we've really broken down the CLI into different optional steps if you want by adding a load of different flags in here. And we've made it more flexible by adding in uh optional like RPC providers. So at hackathons as well as always there's an

122. 54:03

     issue just before deadline. So we had some issues with RPC providers. Now you can just define this this JSON file and use whatever RPC you want inside of the CLI and inside of the the agent as as well. And as part of using it inside the agent, we also shipped version 1.0.0 of shade agentjs. And this is our abstract library to interact and and build shade agents. So instead of know having to know specific uh TE code or

123. 54:44

     specific uh near blockchain code, you can just interact with shade agentjs which exposes a a load of uh nice functions to to do things. So for example here we can request a signature from uh from your shade agent using these NPC accounts and just specify some details there without having to know that there's a there's an agent contract under the hood and there's NPC under the hood. Everything is just done right through this this one function and we expose um a load of different functions

124. 55:29

     and inside this update we have made some breaking changes. So if you're using shade agents already make sure you're using the most recent API image hash and you've updated your CLI to the most recent version. And to um to get into a bit more like more general updates as well, we also um we've built this this shade agent.js which is this library. But it's not just a library. It's actually an API under the hood. And this API is designed so

125. 56:11

     you can build your uh agents in in whatever language you want by just using this API. And you can see here you just need to call the API um which will be like localhost and on port 3000 and then specify which method you want to use.

126. 56:34

     So now you can build an agent in Rust, Python, uh I guess even stranger languages. Maybe you want to build an agent in C, C++. Well, you can do that all now just using this API, not just JavaScript and and TypeScript. We've also added uh support for signing for EDDDSA which now gives support for like Salana, Tron and Sooie. And we had this shade agent template which before was in Nex.js. We've now transformed that to a

127. 57:15

     simple like TypeScript hono server to replicate more how we were going to build an agent in reality. and we have like a locally hosted front end as part of that repo so you can uh interact with your agent a bit a bit easier. But one of the most interesting things we've done here, if you go, we've we've uh we've to custom agent contract, which is really interesting. So if you're trying to do

128. 57:56

     more agent use cases, you can deploy contract and do some cool stuff with that. But the more exciting thing than adding more documentation is actually reducing the amount of documentation. So previously we had this uh we had this sandbox development flow and we also had this production agent development flow. Well because of these updates with the CLI and the API, we've actually managed to merge all of this into just one flow which was called the sandbox. Now we're

129. 58:37

     just calling it the shade agent framework. So if you come into uh yeah docs docs.near.org under AI and agents and shade agents you can go over to the quick start here and learn how to deploy a sample agent in as little steps as as possible. You need some prerequisites here and like how this agent actually works and then uh the key components of it and how you can actually modify the example to fit your own use case. And this is something

130. 59:20

     I was doing yesterday on the um DevHub live stream. You may have seen that where I adapted the template from what is like an Ethereum price oracle into uh this shade agent that operates on Salana to randomly pick an account ID from a list of NFT holders and airdrop them some Salana. So, I'm actually going to switch the tab I'm sharing here. If I can find um give me a minute to pull it up. I think I closed it down

131. 1:00:08

     and I'll share my code tab just over here. All right, let me stop sharing this screen if I can figure it out and share. I There we go. Yeah. So, when you go ahead and clone the template, you're going to see this. Um, I mean this one's a little bit different because we adapted this yesterday to be this Salana example, but when you have this template, you just need to go ahead and uh fill out your environment variables here. And you'll see this docker compose

132. 1:00:56

     file should look quite familiar. You've seen it like two or three times already this stream. And this is what's gets deployed to the the TE. Um, and we have this really nice abstract flow with the shade agent CLI where you can just go ahead and run shade agent CLI. Actually, our run help so we can see the options we have here.

133. 1:01:24

     And we're going to actually do the tag no redeploy because I already have the agent contracts deployed under the hood. So this is going to build our docker image and push it to fa cloud. If you are an old user of shade agents, you will have noticed that already this is operating about 10 times as as quick as it was before. This whole flow used to take about uh five minutes but the app is already live here on far cloud. So if

134. 1:02:00

     I go back to my other screen now back to um let me find the right one. I think we're over here. Yep, there we go. So, uh, yeah, this is this is live on Fal Cloud now. We can see that it's still booting up over here, but just in a minute, we'll be able to go in and see our app is live. You can see it's it's booting up. It usually takes about uh 30 seconds to a minute to build locally with this CLI and then about a minute for it to get up and running on Fowler. And we can see our app is

135. 1:02:53

     running now. So I'm going to try and test this now by copying the endpoint and doing slash a. I'm gonna just actually ping that end point. We can see the app is running and I can do some things like get my agent account ID. Um, that's too many C's.

136. 1:03:23

     So, this is my agent's near account ID and its balance. This is a Salana example. So, I should be able to do Salana account. This still has some balance in from yesterday, which is awesome. So, I can run this example transaction route, which is going to pull up this list of NFT holders from this Madlads collection on Salana, select one of these holders randomly, and uh send them 0.1.

137. 1:03:58

     all. So, this takes a little while to loop through this cuz I think there's like 9,000 NFTts and then we need to rip off a a chain signature, but hopefully we'll get a success result in a minute. Things always go a bit funny with live demos. So, so we'll give it just a few more seconds.

138. 1:04:23

     Aha, it seemed to have worked first time. So you can see I got the recipient address of the random holder and the transaction hash here. This is testn net. So if I go to salana devnet explorer and paste that in here, we should see that 0.1 soul was sent over here.

139. 1:04:52

     So, that's how you get up and started with a shade agent. Um, go to doch.ne.org, check it out, and I hope you I hope you have fun building. Yeah. Owen, are there any upcoming hackathons or anything that people can get involved with just to kind of, you know, get going?

140. 1:05:15

     Yes. So, there um there will be some coming up. Um, we're yet to disclose like when and where they're happening. Uh, but you can expect some happening in the states this year. Um, and I I believe there'll be some activities around like Argentina with Devcon, Dev Connect. So, there's always something happening. And I mean, we're here online all the time. So, just join the shade agent group, ping me if you have any questions, and we also have

141. 1:05:53

     office hours on um on Wednesday's sessions where you can book in for 30 minutes. So, just call come and join us, ask questions, and build something cool. Beautiful. I'd love to see how we can get Visual Diamond integrated somewhere in this development workflow. just make sure people are building secure shade agents.

142. 1:06:18

     Yeah, definitely. And I think uh I'm really excited to an example I'm going to be building soon is actually using Fowlers's uh GPUs so we can actually enable some proper AI use cases inside of shade agents which were one of the reasons they were built. So uh instead of just interacting with a model itself, you'll be able to have like a verified model and then a verified agent that is doing something on top of that model and

143. 1:06:50

     turning that uh AI part into actions which will be really cool. That'd be huge. Awesome. And is there anything else you want to flag before we move on to the final guest? Uh no, that's that's everything from us. just check out the docs and there'll be more coming soon.

144. 1:07:09

     Awesome. Well, Owen, hope to have you have you on soon and uh we're going to move on to Felix. Thanks. See, Felix, welcome back. Hi. Thanks a lot for inviting me again. Really happy to be here. Yeah. Yes. Uh this is going a little bit longer than expected, so appreciate your uh your flexibility. uh if you just don't mind, I know you've been on this show a few times, but for those who don't know you, who are you? What is X-Ray and what have you shipped this week?

145. 1:07:39

     Yeah. Hi everyone. I'm Felix. I'm the CEO and founder of X-Ray. And then we are building the encrypted memory and context for AI agents. So we today we'll show like a 40 end to end encrypted AI agents also integrate with VA so that not only that you get protect from like all the prom and query you have but also all the data is connected to and it will also make sure that the agent when

146. 1:08:03

     you're talking with it with a te all of the data will be you have persistent memory which means that it will recall everything about you based on what you're sharing with it and still you can own all the data because all the data is controlled by your secret key. So I'll show you first about like the architecture that I'm going to demo.

147. 1:08:22

     So this will be the architecture that we're going to demo. Uh any app or like any like for instance notion, slack or any of it that supports our SDK which we support, we'll be able to the user will be able to encrypt and then upload all of their data into this encrypted vector database and your X-ray crowd. As a vector database holder, we can guarantee that we never see your data. And also when the agent is sending query to your data base, we can also guarantee that we

148. 1:08:47

     never see the query and the result but still return the most relevant data to your agent network which really powerful that can unlock a lot of use cases. For instance, like enterprise secure private green where they don't need to require anything to be deployed on prem anymore.

149. 1:09:01

     Now they can have like a fully remoted service but still have the same privacy and security guarantee as on-rem service. And on top of that, we're also working on the encrypted memory. So the agent will actually remember interactions between you and the agent.

150. 1:09:14

     And all of this will be controlled by the user secret key. So only the user will be able to decrypt this uh memory and this data. Yeah. So now I'll quickly show you like how you can work with our system. We created a really developer friendly user dashboard. So how it works is that the developer can create something like a near demo key and then after having this they'll be able to copy it, save it somewhere and then connect to the man uh the agent key that uh the knowledge base

151. 1:09:48

     that I want to work with. So now I can show that I want to connect with the near weekly demo save it and then the developer will be able to track all of the API key usage and also working on like the knowledge base that they have access to and on top of that we created a key that is called the admin key where we can allow AI agent itself to create knowledge base based on their own understanding. So for instance when

152. 1:10:10

     we're chatting with agents it can create a knowledge base about oh this about like our personal chat and it'll create like the private personal chat knowledge base and the work knowledge base and all that. So there's all kind of flexibility and control you can work with on this knowledge base and can not only be controlled by human but also by the agent itself.

153. 1:10:28

     So this is like a really basic proof of concept of how X-rays can work with fala. So this is also the API endpoint of like the large range model is also hosted on the follow server and then we're connecting to the X-face uh vector database. So now you can ask questions about like for instance who am I and then it'll directly have access to like oh to like all your private data and all your memory. So how it works in this demo is that I have already pre-uploaded

154. 1:10:54

     all my uh chbd memory and cloud memory into the system. So now when I'm starting a chat with the follow agent, I directly have access to all my pre directly have access to all my previous uh ideas and how it works here is kind of simple but we can actually make all of this automated as well. So I can upload the data that I want. Let me try to find the place of the I think the data is inside here. So we can have context and then I can put like all the GBT memory here and then you can see

155. 1:11:26

     like the upload procedure is going and after it's uploaded you can see like the data is reflected here. So you can see like previously this is 242 and then now after adding it becomes 248 which means I added six new vectors inside my system and then now you can ask like questions about anything that is personal private or you can also work on connecting to like stack channels notion and all of

156. 1:11:51

     that. So it'll have a persistent memory about you and each user can control their own knowledge graph and then work with any agent that uh they want to work with. Yeah. So that is kind of like a demo and like a showcase of how to build out.

157. 1:12:06

     I mean this is answers a really big issue for me. I am so uncomfortable connecting my Claude or Anthropic account to my Slack to my email to everything. Like I'm just not going to do it. But I really want the power of Claude to essentially you know help me manage my life. And so I'm curious like how do you then plan to take this to sort of next step like you know this exists today how do you get distribution because I think currently a lot of people today are using these you know

158. 1:12:37

     traditional models not traditional but like you know these larger AI company models um how do you plan to integate with them? Yeah. So in the last uh demo like here as well we created like a fully private MCP server that can control all your data from cloud and GPT. So essentially we become like this unified memory layer that can work with any agent that you want to and we can make sure that

159. 1:13:00

     whatever is stored inside our knowledge base only the user have control over it and then they can be the person to choose who they want to share with. For instance, they want for more sensitive data, they can go to follow and get like full confidential cloud or they can have like some specialized small language model on their service or for more like okay data they can also still share it with different uh GBT or all of it and then the point of like allowing so we recreate like this unified memory layer

160. 1:13:25

     and then the reason why people will be willing to do it is because we guarantee we never have access to their data because that's kind of like the mode for all these companies. Yeah, got it. Man, this is so exciting. Um I guess besides follow like how do you plan I guess the MCP server but like do people have to just specifically call your MCP server? How do you plan to like have people discover this?

161. 1:13:47

     Yeah. Yeah. Yeah. So currently we are uh still in a closed source uh stage but we really plan to open source soon. So anyone who kind of filled out like uh this beta testing agreement on our request access on our x-rays.ai AI website we'll give them access to like our SDK so they can kind of use it as like a vector database and then they get like this access to this developer dashboard and then now they can start building interesting agents on top of it

162. 1:14:17

     and then they can also a cool thing about it is that uh if more and more people build upon our Srirax SDK then like all this agent will all support this common memory of the the user so it's no longer like isolated and siloed and then like for developer the good news is that once say a support SDK that get access to the whole user data and they can make measure what user want to share and what they don't want. Yeah. So we're really looking to people trying to

163. 1:14:42

     build on top of us and it'll make it as easy as possible. Yeah. That's amazing. Well any hackathons that proximity is running or near foundation we should absolutely include you guys. Um I love what you're doing especially on all the you know enterprise focused side. Um, I am very skeptical of throwing all my personal information and treating these AIs as therapists and everything like

164. 1:15:04

     that, but I want to. And so, um, yes, I'll be checking on your MCP server as well. Oh, thank you so much. Awesome, Felix. Well, thank you for joining. All right, everybody. Uh, we're about 15 minutes over our normal time, but it was probably because it was jam-packed with exciting, uh, products and so uh, today again is the 20th episode. Uh pretty happy with where the show is going. It's I think every

165. 1:15:28

     episode just getting higher and higher quality. Um thank you all for joining. Again, we do this every single Thursday. Uh next week we already have a couple amazing projects lined up. Uh so if you do want to sort of if you are building in the near AI ecosystem and want to present your work, feel free to DM me on Twitter. I'm Cameron_Dennis.

166. 1:15:46

     Uh I'm quite active and so you'll see me there. Besides that, uh that's all we got. Hope you have a wonderful weekend and hope to see you next week. Thank you. Thanks for watching. Subscribe to Near Week.