{"id":"jQ33aPEw_mU","title":"Zero to Verified: Analyze a Bank Statement With Private AI","description":"A full build walkthrough. We're analyzing a bank statement with private inference on NEAR AI cloud, and proving afterward that nobody along the path could read it or alter the result.\n\nFinancial data is the right test case. You may not want individual transactions exposed, and you almost certainly don't want your whole financial picture sitting with a provider that can train on it. It's also structured data, which is exactly what these models handle well.\n\nHalf this video is about planning, not code. AI can write the code. You still have to do the thinking. The spec sheet is where a build succeeds or falls apart, and I show mine: what the app does, the stack rules, facts pulled straight from the NEAR AI docs, the categories I want transactions sorted into, and the specific verifications I want built in. If you're not sure about your stack, have that conversation with your coding agent before you ask it to build anything.\n\nThe app has four parts. Generate a fake bank statement, or download one. Stand up a local server. Connect to a model. Verify what happened, with a front end that shows the whole workflow so you can watch the server return evidence and the browser display it.\n\nThen the part I care about. Five checks in a proof panel, plus a button that tampers with the response so you can watch a single changed character break the signature. Attestation tells you where the model ran, on hardware that's cryptographically vouched for. The per-response signature tells you what went in and what came out. Together you get both halves, the where and the what, and nobody in the path can edit the result without you finding out.\n\nWorth being precise about what this proves. It proves your data went where it was supposed to and came back untouched. It says nothing about whether the analysis is correct. That still takes your own review or a subject matter expert. Verified is not the same as right.\n\nThe whole run cost about two cents.\n\nChapters\n0:00 What we're building and why\n0:34 The project: analyzing a bank statement\n1:22 How private inference protects the data\n2:19 Inside the TEE, nobody can look\n3:16 Attestation and per-response signatures\n3:47 The four parts of the app\n4:37 Planning matters more than the code\n5:10 Writing a strong spec sheet\n5:50 Walking through my actual spec\n6:39 Stack rules and pulling facts from the docs\n7:17 The project structure it generates\n7:56 Building the verification checks in\n8:37 Nobody can edit results without you knowing\n9:09 What these proofs don't tell you\n9:33 Revisiting the prompt\n10:42 Running it on a real statement\n11:14 The proof panel\n11:27 Tampering with the response on purpose\n12:12 The sample statement it worked from\n12:36 Getting started on NEAR AI cloud\n13:07 Credits, usage, and what this cost\n13:39 Creating an API key with a spend limit\n\nStart at https://near.ai/ and head to the cloud section.\n\nEarlier in this series: what a TEE actually is, and how routing decisions work in code.\n\nCode repo: https://github.com/joe-rlo/nearai-sta...\n\n#NEARAI #PrivateInference #TEE #AIEngineering #BuildInPublic #ConfidentialComputing","channel_name":"NEAR DevHub","published_at":"2026-08-26T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/jQ33aPEw_mU/hqdefault.jpg","transcript_text":"Hey everyone, thanks for joining. Today we're going to do a full walkthrough of how to build with NEAR AI. And what's really neat about this is that we're going to blend leveraging AI to help us code and put together the project, but we're also going to make sure that we walk through the steps of planning. That's a really big piece of all of this. As much as AI can write code for us, we really have to be the ones to do the planning for us to get the results that we want to get. Let's walk through this project that we're going to do today using NEAR AI and focusing on privacy. I'm calling this Zero to Verified. And what we're going to do today, instead of building another chatbot or something like that, we're going to analyze a bank statement with private inference on NEAR AI Cloud. Now, this is a great example because a lot of times financial information obviously you want to be private. There might be things in there that not only are just private transactions, but as a whole, you just maybe don't want your financial information leaked to another provider. So we're going to leverage the ability to have private inference on this information. Also, this is a great use case for AI to go in and do analysis over some structured data which we can get from a bank statement. So let's get started. Let me set the scene a little bit for you here. When you are dealing with sensitive information, sensitive data, the key that we need to focus on is it needs to be able to have privacy in both the user having their data protected, just like how it's being stored and how it's being transferred, but also then what happens when it goes through AI inference. And NEAR AI, because it is using TEEs, which is from another video that we can link back to so you can see how TEEs work. But trusted execution environments allow us to be able to have privacy for these types of inference, which you don't normally get from an OpenAI or an Anthropic or Google, who might take that information and use it for their own training data. If we look at this, we see, okay, we're going to do a sample thing. We're going to look at some different statements here and see if we can figure out what's happening with our transactions. And the way the model is going to run is that it goes in a way where nobody can look. So if you've already seen my video on TEEs, trusted execution environments, then you would know that basically it is a private room that the data goes into that nobody can peek in. So what's happening here is that your machine would send a file over TLS, which is again standard for encryption, and then the model, which is sitting inside the TEE, can then process that. The cloud provider, whoever the actual model provider is doing that, and also NEAR AI cannot read that. So the model provider can't read it, the cloud provider can't read it, no one can read it. It is fully private. So, you know, that's secure, and we're going to get to that too, of how we can prove that. So keep that thought in mind. How can you prove it? So we have something called attestation. Attestation is the enclave that produces that cryptographically signed piece of evidence that says this hardware is secure, this model is running inside this secure hardware, so on and so forth. That's the proof of where it ran. Then we want to have the per-response signature, which is going to prove what went in and then what went out, which is now you have both sides of this, of the where and the what. What are we going to build in this example? This example is going to be fairly straightforward, but we're going to have this in kind of four key parts. We're going to generate a fake bank statement. So that's easy, but you can also download one from the internet or whatever. We're going to set up a sample localhost server that will do all of this. We're going to connect to a model, and from there we're going to make sure that we're verifying everything that actually happened on that model. And we're going to have a front end that is going to run and see this. We can actually visualize how this all works. That way the server can return evidence, the browser can show you the correct thing, and you can see this full workflow. Now, here is something that is important. This is really where the lesson begins on how to actually build with AI. We need to start to think about what actually is going to be built. And you can do this two different ways. One, if you feel confident in your ability to decide a tech stack and to talk about it, you can go and define that. If you're not comfortable, these are things that you can have conversations about with your AI system that's going to do the building beforehand, and figure out what might be the best fit for what you're trying to build. Now, the key is though, if you're going to do that, or in any form, you have to have a strong specification. That spec sheet is going to really define how this is actually going to work, and is this going to be built correctly. Okay, so let's switch to our spec sheet and go from there. Depending on what AI you're building with, whether it's Codex or Claude Code or whatever other system maybe you're using for this, doesn't really matter. If you're using Claude Code, you usually use a CLAUDE file that will have your specs, but you just need a spec sheet. So it doesn't have to be a CLAUDE.md. And here we're first going to define, and again your AI can do this. So you can have a conversation using one of the chatbots and they can help write the spec sheet for you. So don't feel like you have to write this all by yourself. But again, you do need to know what you're trying to build and be very clear about that. Here's what I wrote. So it's a tutorial app. A user is going to upload a bank statement, and the server is going to extract the text, send it to a model inside of a TEE, and return the structured JSON and categorize everything inside there. And we want to make sure that we validate that it went through the TEE and everything was not altered. And the idea here is we want to show a proof panel, and we don't want it to be a chatbot, because then we're going to get into more detail as we go, typically. Then again, everyone's a little bit different. So this is just my way of doing it, but you can do it your own way. So we want to have some rules about what kind of stack we're using. And again, this can be the conversation that you have with your agent in trying to build this, or you can write it yourself by hand. We would then also want to have some facts, and this usually can come from documentation directly. So in this example here, for the NEAR AI Cloud facts, I actually shared the docs from NEAR AI to Claude. Then from there, it breaks all that down, including the extraction, all of that stuff. I can then say, okay, what do we want to actually verify. I put in some things I wanted to verify and go from there. Now once it has its specs and you tell it to build, you should get a stack, if you're using this example directly, something very similar to what I have here on the right hand side. So you might have something that resembles this layout, which is a standard kind of Node project. Here's my server, here's my web front end, here's some scripts, docs, things like that, and put it all together depending on what you're doing. If you're doing something in React, if you're doing something in Python, it's going to be structured a little bit differently. But again, these AI systems are advanced enough that they're able to go and determine this for you, and they know what to do. The check part was also really important for this project, because this is a demo. But these are things that we also want to always highlight whenever we are building with privacy. Even if you don't want to ship it to the customer, during the development phase it's always good to have some kind of verification process, that what you were doing and what you were testing is actually verifiable. And so we broke this down, and you'll see when I show you the demo that there are five checks here that we can look at, and then we can even alter and check the validation correctly. How does it look when it's tampered with by changing that? And so again, really simple, but it's important to understand that nobody in this path can edit the results without you finding out, if you are actually doing these checks. That's the important piece of building with privacy and building with something that is on top of a TEE, is that you can always have checks, and the end user can even check independently. It doesn't have to be you as the developer. So this is a really important aspect of building with privacy, that you can have this part as part of it. It does highlight for you that no one could read it, the exact data was sent correctly. So again, it's important to always remember that there is no rightness in any of these proofs. Rightness comes from separate verification that you may have to do from a subject matter expert, or doing your own checks, or depending on what you're doing, that has to maybe happen separately. So let's check out what this built and how this all works. So again, let me remind you how simple overall this is in the prompt to do it. So if we pull up the prompt again, starting from here of what it was, what I want the stack to be, some different notes and rules around it, some different information about the NEAR AI Cloud and how it's supposed to work. Here's the NVIDIA verification. Here's how we want the data structured. Again, you can have AI determine this. If you don't know it, if you know what you want or how you want it to be categorized, you can do that. So if I know I want these, these are the categories I want in my transaction analyzer. Great. I can write these all in myself, and then what verifications I want to have. I can give all of this to Codex, to Claude Code, whatever it might be, to go build this, and then have something here. It is really simple. And I'm going to add in a bank statement to show you what it looks like when this actually runs, and the attestation part and all of that. Okay, here's my document. It's now going to extract and analyze. And here's our results. So if we look through these results here, it's going to flag some different information about the data. It gave me some analysis of what's going on here, some things for me to look at. It broke down my spending. It broke down the merchants, all of that. We can see all the transactions. All the transactions are here, but we'll go collapse that. And then here's my proof of private inference. So this is the part that I wanted to put in, to highlight what we're doing here with NEAR AI and private clouds. And I can see, okay, all of this has passed. Here's what it's been signed with. Here are the different proofs. Now I can simulate tampering with the response. So the response of what that's supposed to actually be checked with, I can tamper with it with this test button we put in. And I see, okay, one character changed, signature no longer matches. So now we say, oh, this doesn't match. This is why having these checks, even if they're silently in the background as something that you're building, is so important. If someone were to modify these results, someone were to be a man in the middle of some sort or whatever, you have that, and it gives peace of mind to your users who are expecting privacy all the way through. Uh-oh, something's wrong. You know, then that can be addressed. Restore it back and you can see everything matches up accordingly. To get a sense of what I used, here's a sample bank statement. You can see this is what I gave it. It broke all this data down. If you've used AI in any form, you probably see that it does really well with these types of documents. This is just a PDF. It can be whatever. But again, that's why we're leveraging the AI to do all of this and to do that work, and that's why we want that private inference to process this information. You can start today at near.ai. Go to the cloud section. You can register and sign up. Let me take a second here before we wrap up and show you how to easily get started with that. I'm on the website. I can go down to Your Cloud, start building. I'm already logged in. You can just sign in at any time. You can use your email. There's all different ways that you can sign in with it. So just sign in, create an account. You can put money in to actually have credits there. You can also see here that you can stake NEAR, the NEAR token, if you want to do that to earn credits as well. If I go to My Usage, this is the usage that we just did. I used a newer model, Qwen 3.5. You can see how many tokens were just used. Cost me about 2 cents to do that. So you don't have to put a lot of money in to get started. You can just put in a little bit of money. And now my keys. I can go to Keys. I can hit create a new key. I can also put a limit on how much that key can spend as well, which is a really nice option. And then just name that, and it gives you a key that you can then get started and build with. If you missed any of the previous videos on how to build with AI, how to use private inference and routing and things like that, or what a TEE actually is, I encourage you to check out those previous videos, and we'll link to those as well. Until next time, I'm Joe, and I hope you learned something. Thanks.","transcript_synced_at":"2026-11-03T03:47:12.818Z","segments":[{"start_seconds":0,"text":"Hey everyone, thanks for joining. Today we're going to do a full walkthrough of how to build with NEAR AI. And what's really neat about","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=0s"},{"start_seconds":10,"text":"this is that we're going to blend leveraging AI to help us code and put together the project, but we're also going to make sure that we walk through the steps of planning. That's a really big piece of all of this.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=10s"},{"start_seconds":21,"text":"As much as AI can write code for us, we really have to be the ones to do the planning for us to get the results that we want to get.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=21s"},{"start_seconds":28,"text":"Let's walk through this project that we're going to do today using NEAR AI and focusing on privacy.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=28s"},{"start_seconds":34,"text":"I'm calling this Zero to Verified. And what we're going to do today, instead of building another chatbot or something like that, we're going to analyze a bank statement with private inference on NEAR AI Cloud.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=34s"},{"start_seconds":46,"text":"Now, this is a great example because a lot of times financial information obviously you want to be private.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=46s"},{"start_seconds":52,"text":"There might be things in there that not only are just private transactions, but as a whole, you just maybe don't want your financial information leaked to another provider.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=52s"},{"start_seconds":63,"text":"So we're going to leverage the ability to have private inference on this information. Also, this is a great","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=63s"},{"start_seconds":69,"text":"use case for AI to go in and do analysis over some structured data which we can get from a bank statement. So","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=69s"},{"start_seconds":77,"text":"let's get started. Let me set the scene a little bit for you here.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=77s"},{"start_seconds":82,"text":"When you are dealing with sensitive information, sensitive data, the key that we need to focus on is it needs to","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=82s"},{"start_seconds":91,"text":"be able to have privacy in both the user having their data protected, just like how it's being stored and how it's being","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=91s"},{"start_seconds":100,"text":"transferred, but also then what happens when it goes through AI inference. And NEAR AI, because it is using","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=100s"},{"start_seconds":107,"text":"TEEs, which is from another video that we can link back to so you can see how TEEs work. But trusted execution","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=107s"},{"start_seconds":115,"text":"environments allow us to be able to have privacy for these types of inference, which you don't normally get from an","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=115s"},{"start_seconds":123,"text":"OpenAI or an Anthropic or Google, who might take that information and use it for their own training data. If we look at this, we see, okay, we're going to do","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=123s"},{"start_seconds":131,"text":"a sample thing. We're going to look at some different statements here and see if we can figure out what's happening with our transactions.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=131s"},{"start_seconds":139,"text":"And the way the model is going to run is that it goes in a way where nobody can look. So if you've already seen my","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=139s"},{"start_seconds":147,"text":"video on TEEs, trusted execution environments, then you would know that basically it is a private room that the data goes into that nobody can peek in.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=147s"},{"start_seconds":158,"text":"So what's happening here is that your machine would send a file over TLS, which is again standard for encryption,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=158s"},{"start_seconds":166,"text":"and then the model, which is sitting inside the TEE, can then process that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=166s"},{"start_seconds":173,"text":"The cloud provider, whoever the actual model provider is doing that, and also NEAR AI cannot read that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=173s"},{"start_seconds":182,"text":"So the model provider can't read it, the cloud provider can't read it, no one can read it. It is fully private. So, you know,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=182s"},{"start_seconds":189,"text":"that's secure, and we're going to get to that too, of how we can prove that. So keep that thought in mind. How can you prove it? So we have something called attestation.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=189s"},{"start_seconds":200,"text":"Attestation is the enclave that produces that cryptographically signed piece of evidence that says this hardware","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=200s"},{"start_seconds":210,"text":"is secure, this model is running inside this secure hardware, so on and so forth. That's the proof of where it ran.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=210s"},{"start_seconds":218,"text":"Then we want to have the per-response signature, which is going to prove what went in and then what went out, which is now you have both sides of this, of the where and the what.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=218s"},{"start_seconds":230,"text":"What are we going to build in this example? This example is going to be fairly straightforward, but we're going to have this in kind of four key parts. We're going to generate a fake","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=230s"},{"start_seconds":240,"text":"bank statement. So that's easy, but you can also download one from the internet or whatever. We're going to set up a","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=240s"},{"start_seconds":248,"text":"sample localhost server that will do all of this. We're going to connect to a model, and from there we're going","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=248s"},{"start_seconds":255,"text":"to make sure that we're verifying everything that actually happened on that model. And we're going to have a front end that is going to run and see this.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=255s"},{"start_seconds":264,"text":"We can actually visualize how this all works. That way the server can return evidence, the browser can show you the correct thing, and you can see this full workflow. Now, here is something that is important.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=264s"},{"start_seconds":277,"text":"This is really where the lesson begins on how to actually build with AI.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=277s"},{"start_seconds":283,"text":"We need to start to think about what actually is going to be built. And you can do this two different ways.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=283s"},{"start_seconds":289,"text":"One, if you feel confident in your ability to decide a tech stack and to talk about it, you can go and define that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=289s"},{"start_seconds":296,"text":"If you're not comfortable, these are things that you can have conversations about with your AI system that's going to do the building beforehand,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=296s"},{"start_seconds":304,"text":"and figure out what might be the best fit for what you're trying to build. Now, the key is though, if you're going to do that, or in any form,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=304s"},{"start_seconds":313,"text":"you have to have a strong specification. That spec sheet is going to really define how this is actually going to work,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=313s"},{"start_seconds":321,"text":"and is this going to be built correctly. Okay, so let's switch to our spec sheet and go from there.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=321s"},{"start_seconds":327,"text":"Depending on what AI you're building with, whether it's Codex or Claude Code or whatever other system maybe you're using for this, doesn't really matter.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=327s"},{"start_seconds":337,"text":"If you're using Claude Code, you usually use a CLAUDE file that will have your specs, but you just need a spec sheet.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=337s"},{"start_seconds":341,"text":"So it doesn't have to be a CLAUDE.md. And here we're first going to define, and again your AI can do this.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=341s"},{"start_seconds":350,"text":"So you can have a conversation using one of the chatbots and they can help write the spec sheet for you. So don't feel like you have to write this","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=350s"},{"start_seconds":357,"text":"all by yourself. But again, you do need to know what you're trying to build and be very clear about that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=357s"},{"start_seconds":362,"text":"Here's what I wrote. So it's a tutorial app. A user is going to upload a bank statement, and the server is going to extract the text,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=362s"},{"start_seconds":369,"text":"send it to a model inside of a TEE, and return the structured JSON and categorize everything inside there. And we want to make sure that we validate that it went through the TEE and everything was not altered.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=369s"},{"start_seconds":381,"text":"And the idea here is we want to show a proof panel, and we don't want it to be a chatbot, because then we're going to get into more detail as we go, typically. Then again, everyone's a little bit different.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=381s"},{"start_seconds":393,"text":"So this is just my way of doing it, but you can do it your own way. So we want to have some rules about what kind of stack we're using. And again, this can be the conversation","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=393s"},{"start_seconds":402,"text":"that you have with your agent in trying to build this, or you can write it yourself by hand. We would then also want to have some facts, and this usually can come from documentation","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=402s"},{"start_seconds":413,"text":"directly. So in this example here, for the NEAR AI Cloud facts, I actually shared the docs from NEAR AI to Claude.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=413s"},{"start_seconds":424,"text":"Then from there, it breaks all that down, including the extraction, all","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=424s"},{"start_seconds":430,"text":"of that stuff. I can then say, okay, what do we want to actually verify. I put in some things I wanted to verify and go from there.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=430s"},{"start_seconds":440,"text":"Now once it has its specs and you tell it to build, you should get a stack, if you're using this example","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=440s"},{"start_seconds":446,"text":"directly, something very similar to what I have here on the right hand side. So you might have something that resembles this layout, which is a standard kind of Node project.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=446s"},{"start_seconds":457,"text":"Here's my server, here's my web front end, here's some scripts, docs, things like that, and put it all together depending on what you're doing. If you're doing something in React, if you're doing something in","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=457s"},{"start_seconds":466,"text":"Python, it's going to be structured a little bit differently. But again, these AI systems are advanced enough that they're able to go and determine this for you, and they know","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=466s"},{"start_seconds":476,"text":"what to do. The check part was also really important for this project, because this is a demo. But these are things that we also want to always highlight whenever we are building with privacy.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=476s"},{"start_seconds":485,"text":"Even if you don't want to ship it to the customer, during the development phase it's always good to have some kind of verification process, that what you were doing and what you were testing is actually verifiable.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=485s"},{"start_seconds":497,"text":"And so we broke this down, and you'll see when I show you the demo that there are five checks here that we can","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=497s"},{"start_seconds":504,"text":"look at, and then we can even alter and check the validation correctly. How does it look when it's tampered with by","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=504s"},{"start_seconds":513,"text":"changing that? And so again, really simple, but it's important to understand that nobody in this path can edit","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=513s"},{"start_seconds":519,"text":"the results without you finding out, if you are actually doing these checks. That's the important piece of building with privacy and building with something","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=519s"},{"start_seconds":527,"text":"that is on top of a TEE, is that you can always have checks, and the end user can even check independently. It doesn't have to be you as the developer.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=527s"},{"start_seconds":536,"text":"So this is a really important aspect of building with privacy, that you can have this part as","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=536s"},{"start_seconds":542,"text":"part of it. It does highlight for you that no one could read it, the exact data was","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=542s"},{"start_seconds":549,"text":"sent correctly. So again, it's important to always remember that there is no rightness in any of these","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=549s"},{"start_seconds":557,"text":"proofs. Rightness comes from separate verification that you may have to do from a subject matter expert, or doing your own checks, or depending on what you're doing, that has to maybe happen separately.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=557s"},{"start_seconds":568,"text":"So let's check out what this built and how this all works.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=568s"},{"start_seconds":573,"text":"So again, let me remind you how simple overall this is in the prompt to","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=573s"},{"start_seconds":580,"text":"do it. So if we pull up the prompt again, starting from here of what it was,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=580s"},{"start_seconds":589,"text":"what I want the stack to be, some different notes and rules around it, some different information about the NEAR AI Cloud and how it's supposed to","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=589s"},{"start_seconds":598,"text":"work. Here's the NVIDIA verification. Here's how we want the data structured. Again, you can have AI","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=598s"},{"start_seconds":606,"text":"determine this. If you don't know it, if you know what you want or how you want it to be categorized, you can do that. So if I know I want these, these are the categories I want in my transaction analyzer. Great.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=606s"},{"start_seconds":617,"text":"I can write these all in myself, and then what verifications I want to have. I can give all of this to Codex, to Claude Code, whatever it might be,","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=617s"},{"start_seconds":626,"text":"to go build this, and then have something here. It is really simple. And I'm going to add in a bank statement to show","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=626s"},{"start_seconds":635,"text":"you what it looks like when this actually runs, and the attestation part and all of that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=635s"},{"start_seconds":642,"text":"Okay, here's my document. It's now going to extract and analyze.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=642s"},{"start_seconds":650,"text":"And here's our results. So if we look through these results here, it's going to flag some different","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=650s"},{"start_seconds":657,"text":"information about the data. It gave me some analysis of what's going on here, some things for me to look at. It broke down my spending.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=657s"},{"start_seconds":669,"text":"It broke down the merchants, all of that. We can see all the transactions. All the transactions are here, but we'll go collapse that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=669s"},{"start_seconds":676,"text":"And then here's my proof of private inference.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=676s"},{"start_seconds":679,"text":"So this is the part that I wanted to put in, to highlight what we're doing here with NEAR AI and private clouds. And I can see, okay, all of this has passed. Here's what it's been signed with.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=679s"},{"start_seconds":690,"text":"Here are the different proofs. Now I can simulate tampering with the response. So the response of what that's supposed to actually be checked with, I can tamper with it with this test button we put in.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=690s"},{"start_seconds":701,"text":"And I see, okay, one character changed, signature no longer matches.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=701s"},{"start_seconds":706,"text":"So now we say, oh, this doesn't match. This is why having these checks, even if they're silently in the background as something that you're","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=706s"},{"start_seconds":713,"text":"building, is so important. If someone were to modify these results, someone were to be a man in the middle of some sort or whatever, you have that, and","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=713s"},{"start_seconds":721,"text":"it gives peace of mind to your users who are expecting privacy all the way through. Uh-oh, something's wrong.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=721s"},{"start_seconds":729,"text":"You know, then that can be addressed. Restore it back and you can see everything matches up accordingly.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=729s"},{"start_seconds":737,"text":"To get a sense of what I used, here's a sample bank statement. You can see this is what I gave it.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=737s"},{"start_seconds":744,"text":"It broke all this data down. If you've used AI in any form, you probably see that it does really well with these types of documents. This is just a PDF. It can be whatever.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=744s"},{"start_seconds":752,"text":"But again, that's why we're leveraging the AI to do all of this and to do that work, and that's why we want that private inference to process this information. You can start today at near.ai.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=752s"},{"start_seconds":762,"text":"Go to the cloud section. You can register and sign up. Let me take a second here before we wrap up and show you how to easily get started with that.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=762s"},{"start_seconds":770,"text":"I'm on the website. I can go down to Your Cloud, start building. I'm already logged in. You can just sign in at any time. You can use your email.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=770s"},{"start_seconds":780,"text":"There's all different ways that you can sign in with it.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=780s"},{"start_seconds":783,"text":"So just sign in, create an account. You can put money in to actually have credits there. You can also see here that you can stake NEAR, the NEAR token, if you want to do that to earn credits as well.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=783s"},{"start_seconds":798,"text":"If I go to My Usage, this is the usage that we just did. I used a newer model, Qwen 3.5.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=798s"},{"start_seconds":808,"text":"You can see how many tokens were just used. Cost me about 2 cents to do that. So you don't have to put a lot of money in to get started.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=808s"},{"start_seconds":816,"text":"You can just put in a little bit of money. And now my keys. I can go to Keys. I can hit create a new key.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=816s"},{"start_seconds":824,"text":"I can also put a limit on how much that key can spend as well, which is a really nice option. And then just name that, and it gives you a key that you can","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=824s"},{"start_seconds":833,"text":"then get started and build with. If you missed any of the previous videos on how to build with AI, how to use private inference and routing and things like","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=833s"},{"start_seconds":841,"text":"that, or what a TEE actually is, I encourage you to check out those previous videos, and we'll link to those as well. Until next time, I'm Joe, and I hope you learned something. Thanks.","url":"https://www.youtube.com/watch?v=jQ33aPEw_mU&t=841s"}],"youtube_url":"https://youtube.com/watch?v=jQ33aPEw_mU","url":"https://near-pedia.xyz/videos/jQ33aPEw_mU"}