WEBVTT 00:00:00.000 --> 00:00:02.500 Imagine this. You send a quick message, maybe 00:00:02.500 --> 00:00:05.740 on Telegram, maybe Slack, and just seconds later, 00:00:05.879 --> 00:00:09.699 you get back a really detailed professional technical 00:00:09.699 --> 00:00:12.779 stock analysis. Yeah, like full breakdown, mass 00:00:12.779 --> 00:00:16.179 CD, support, resistance lines, even the actual 00:00:16.179 --> 00:00:18.559 chart image right there on your phone. Exactly. 00:00:18.679 --> 00:00:19.879 It sounds like something you'd need, I don't 00:00:19.879 --> 00:00:22.199 know, expensive financial software for. Right. 00:00:22.579 --> 00:00:25.559 But here's the kicker. This whole thing, this 00:00:25.559 --> 00:00:27.579 advanced agent system, it was built in just a 00:00:27.579 --> 00:00:31.059 few hours. And get this completely no code. Wow. 00:00:31.300 --> 00:00:33.399 Yeah, it's not just some simple chatbot fetching 00:00:33.399 --> 00:00:37.240 a price. It's a proper AI agent. It's even running 00:00:37.240 --> 00:00:39.920 a pretty sophisticated dual model strategy. Welcome 00:00:39.920 --> 00:00:42.759 to the deep dive. So today we're going to open 00:00:42.759 --> 00:00:44.719 up the hood on that exact system, the personal 00:00:44.719 --> 00:00:47.780 AI stock analyst. Our mission really is to get 00:00:47.780 --> 00:00:49.899 our heads around the architecture. This two part. 00:00:50.350 --> 00:00:52.530 two AI models set up that makes it all tick. 00:00:52.729 --> 00:00:54.770 Yeah, we'll map out what it can actually do, 00:00:54.829 --> 00:00:56.810 you know, memory comparisons, that kind of stuff. 00:00:56.909 --> 00:00:59.630 Then we'll really dig into that core idea, the 00:00:59.630 --> 00:01:03.590 brain versus the specialist separation. And crucially, 00:01:03.670 --> 00:01:06.349 you'll see why they picked two totally different 00:01:06.349 --> 00:01:09.010 AI models. You know, one that's great for reasoning, 00:01:09.129 --> 00:01:11.769 the thinking part, and another one completely 00:01:11.769 --> 00:01:14.549 separate that's amazing at visual stuff, like 00:01:14.549 --> 00:01:16.530 reading those charts. All right, let's dive in. 00:01:16.609 --> 00:01:19.799 Let's do it. okay so let's unpack what this system 00:01:19.799 --> 00:01:23.120 actually does day to day it seems to handle three 00:01:23.120 --> 00:01:25.700 main things that make it feel like a proper agent 00:01:25.700 --> 00:01:28.840 not just you know a simple script right the first 00:01:28.840 --> 00:01:31.060 one's pretty straightforward single stock analysis 00:01:31.060 --> 00:01:35.129 you feed it a ticker symbol like uh NASDAQ .UXU 00:01:35.129 --> 00:01:37.170 or whatever. Exactly. And you get back a full 00:01:37.170 --> 00:01:40.329 professional style report. Trend health, key 00:01:40.329 --> 00:01:43.689 price levels, mass D analysis, volume. Yeah. 00:01:43.750 --> 00:01:45.730 The whole nine yards. That's kind of the baseline. 00:01:45.829 --> 00:01:47.730 Yeah. Yeah. But the second function, comparative 00:01:47.730 --> 00:01:50.370 analysis, that's where this split architecture 00:01:50.370 --> 00:01:53.170 really starts to shine, I think. How so? Well, 00:01:53.209 --> 00:01:55.989 you can ask it to compare two stocks, say Alphabet 00:01:55.989 --> 00:01:57.969 versus Microsoft. Yeah. Like in the example, 00:01:58.109 --> 00:02:00.250 it doesn't just give you two separate reports. 00:02:00.879 --> 00:02:03.500 It runs the entire technical breakdown for both, 00:02:03.659 --> 00:02:05.920 generates the charts for each one, and then, 00:02:05.980 --> 00:02:09.120 this is the cool part, it performs this higher 00:02:09.120 --> 00:02:11.900 level reasoning to give you a final summary comparing 00:02:11.900 --> 00:02:14.879 them side by side. Ah, so it synthesizes the 00:02:14.879 --> 00:02:17.599 information. It doesn't just fetch it. Precisely. 00:02:17.840 --> 00:02:20.379 And that leads to the third really key function, 00:02:20.560 --> 00:02:22.719 smart memory. Right. The conversational aspect. 00:02:22.979 --> 00:02:25.319 Yeah. So if you just ask for that alphabet analysis, 00:02:25.680 --> 00:02:27.340 you can immediately follow up with something 00:02:27.340 --> 00:02:29.500 simple like, OK, now compare that to Microsoft. 00:02:29.500 --> 00:02:32.520 And it knows what that. refers to. Exactly. It 00:02:32.520 --> 00:02:34.419 remembers the last, say, five messages in the 00:02:34.419 --> 00:02:36.740 conversation. So you don't have to repeat the 00:02:36.740 --> 00:02:41.000 ticker. You don't waste time or API credits reanalyzing 00:02:41.000 --> 00:02:43.139 the first doc. It just picks up where you left 00:02:43.139 --> 00:02:45.180 off. That makes sense. And I guess that complexity 00:02:45.180 --> 00:02:47.539 needing memory, needing to synthesize comparisons, 00:02:47.919 --> 00:02:50.960 that's what drives this dual architecture idea. 00:02:51.240 --> 00:02:53.340 You got it. Separating the manager from the doer. 00:02:53.699 --> 00:02:57.400 So part one is the AI agent brain. Think of it 00:02:57.400 --> 00:03:00.039 like the CEO. This is the main workflow. probably 00:03:00.039 --> 00:03:03.439 built in something like NADN. It handles the 00:03:03.439 --> 00:03:05.560 chat, keeps track of the conversation history, 00:03:05.719 --> 00:03:08.219 that memory decides when it needs help from a 00:03:08.219 --> 00:03:10.699 specialist, and then takes the specialist output 00:03:10.699 --> 00:03:13.620 and formats the nice final report for you. And 00:03:13.620 --> 00:03:16.759 part two is? Part two, the technical analysis 00:03:16.759 --> 00:03:20.120 tool. The specialist. This is a totally separate, 00:03:20.219 --> 00:03:23.199 smaller workflow. A sub -workflow. Okay, a sub 00:03:23.199 --> 00:03:25.419 -workflow. Let's define that quickly. Sure. A 00:03:25.419 --> 00:03:27.599 sub workflow is basically like a dedicated mini 00:03:27.599 --> 00:03:29.800 program the main system can call up whenever 00:03:29.800 --> 00:03:32.120 it needs a specific task done. Like calling a 00:03:32.120 --> 00:03:34.979 consultant for a specific job. Exactly like that. 00:03:35.080 --> 00:03:37.719 This specialist one, it does one thing. Perfectly. 00:03:37.840 --> 00:03:40.379 Takes a stock ticker, calls an outside service 00:03:40.379 --> 00:03:42.860 and API to generate the chart image, and then 00:03:42.860 --> 00:03:45.000 does the raw technical analysis on that chart. 00:03:45.159 --> 00:03:48.039 OK, but I have to ask, doesn't splitting it into 00:03:48.039 --> 00:03:50.800 a CEO and a specialist make things more complex, 00:03:51.020 --> 00:03:53.759 like harder to build and manage, especially in 00:03:53.759 --> 00:03:56.500 a no code setup? Why not just cram it all into 00:03:56.500 --> 00:03:59.360 one big workflow? That's a great question. It 00:03:59.360 --> 00:04:01.419 seems counterintuitive, right? But actually, 00:04:01.520 --> 00:04:03.500 it makes things way less complex in the long 00:04:03.500 --> 00:04:06.219 run. Really? By separating them. The main CEO 00:04:06.219 --> 00:04:08.500 system stays really clean and focused just on 00:04:08.500 --> 00:04:11.099 managing the conversation and logic. It's modular. 00:04:11.460 --> 00:04:13.840 So let's say tomorrow you want to add a tool 00:04:13.840 --> 00:04:17.420 that analyzes, I don't know, options data. You 00:04:17.420 --> 00:04:20.019 just build a new separate specialist sub -workflow 00:04:20.019 --> 00:04:22.860 for options, tell the CEO about it, and plug 00:04:22.860 --> 00:04:25.439 it in. The core brain doesn't get cluttered. 00:04:25.459 --> 00:04:27.660 It makes the whole system much easier to manage 00:04:27.660 --> 00:04:30.750 and scale up with new tools later. Ah, I see. 00:04:30.850 --> 00:04:33.670 So that separation actually promotes clarity 00:04:33.670 --> 00:04:36.389 and makes it easier to add more capabilities 00:04:36.389 --> 00:04:39.509 down the road. Precisely. Scalable clarity. That's 00:04:39.509 --> 00:04:41.689 the goal. Okay, so we've got the CEO of the brain. 00:04:41.889 --> 00:04:44.730 Now, which AI model do you actually pick for 00:04:44.730 --> 00:04:46.889 that role? It used to be, you know, you'd just 00:04:46.889 --> 00:04:48.730 go with the biggest name or whatever was hyped 00:04:48.730 --> 00:04:50.629 that week. Right, whatever was trending on Twitter. 00:04:50.910 --> 00:04:52.750 Exactly. Yeah. But for something like financial 00:04:52.750 --> 00:04:54.990 analysis, you need proof it actually performs 00:04:54.990 --> 00:04:57.220 under pressure. The source material mentions 00:04:57.220 --> 00:05:00.579 this platform, mouse1 .ai, that runs something 00:05:00.579 --> 00:05:03.420 called the Alpha Arena. The Alpha Arena. What 00:05:03.420 --> 00:05:05.579 is that exactly? Sounds intense. It kind of is. 00:05:05.720 --> 00:05:08.779 It's basically this live high stakes competition 00:05:08.779 --> 00:05:11.699 for different AI models. They compete in this 00:05:11.699 --> 00:05:14.980 real time streaming environment designed to simulate, 00:05:15.079 --> 00:05:18.220 you know, dynamic market conditions. So it tests 00:05:18.220 --> 00:05:21.100 how well they reason and how reliable they are 00:05:21.100 --> 00:05:22.860 when things are constantly changing, like in 00:05:22.860 --> 00:05:25.529 actual trading. You got it. It's about battle 00:05:25.529 --> 00:05:27.990 -tested performance, not just benchmark scores 00:05:27.990 --> 00:05:30.750 or, you know, how fast it spits out tokens. Okay, 00:05:30.810 --> 00:05:33.649 so based on that kind of testing, which model 00:05:33.649 --> 00:05:36.490 did they choose for the brain? The data apparently 00:05:36.490 --> 00:05:41.189 pointed towards using DeepSeq v3 .1 chat. That 00:05:41.189 --> 00:05:43.589 became the core reasoning engine, the CEO brain. 00:05:43.870 --> 00:05:45.889 DeepSeq. Okay. And how do you actually connect 00:05:45.889 --> 00:05:48.029 to it? They used OpenRouter. Right, OpenRouter. 00:05:48.069 --> 00:05:50.329 Let's quickly define that too. Yeah, simple one. 00:05:50.629 --> 00:05:52.310 OpenRouter is just an aggregator that lets you 00:05:52.310 --> 00:05:54.769 access lots of different AI models using just 00:05:54.769 --> 00:05:57.670 one API key. Super flexible. Gives you options 00:05:57.670 --> 00:06:00.389 if you want to swap models later. Exactly. Now, 00:06:00.509 --> 00:06:02.850 critical for this brain, remember that smart 00:06:02.850 --> 00:06:04.449 memory we talked about? Yeah, remembering the 00:06:04.449 --> 00:06:07.600 last five messages. Right. The absolute crucial 00:06:07.600 --> 00:06:10.100 setup step there, and this is the bit people 00:06:10.100 --> 00:06:12.399 apparently mess up all the time, is you must 00:06:12.399 --> 00:06:15.819 tie that memory to the specific chat ID coming 00:06:15.819 --> 00:06:18.399 from the Telegram trigger. Ah, okay. So the memory 00:06:18.399 --> 00:06:21.300 is unique to each user conversation. What happens 00:06:21.300 --> 00:06:23.459 if you forget that link? What's the failure mode? 00:06:23.680 --> 00:06:27.180 Oh, it's instant chaos. Utter chaos. Imagine 00:06:27.180 --> 00:06:30.459 you're asking about Apple, right? And someone 00:06:30.459 --> 00:06:33.120 else in a totally separate chat just asked about 00:06:33.120 --> 00:06:36.240 Tesla. Uh -oh. Yeah. Suddenly you might start 00:06:36.240 --> 00:06:38.199 getting the Tesla analysis popping up in your 00:06:38.199 --> 00:06:40.740 Apple chat. The conversation memory gets completely 00:06:40.740 --> 00:06:43.019 crossed. It just obliterates the user experience. 00:06:43.259 --> 00:06:45.259 You have to link it to that unique chat ID. Got 00:06:45.259 --> 00:06:47.420 it. Okay. So you've picked the model based on 00:06:47.420 --> 00:06:49.060 performance. You've set up the memory correctly. 00:06:49.319 --> 00:06:51.720 Right. What's next? Defining what the agent should 00:06:51.720 --> 00:06:53.620 actually do. Exactly. You need to give it its 00:06:53.620 --> 00:06:55.899 job description, its personality, its rules, 00:06:56.060 --> 00:06:58.600 and that's done via the system prompt. The prompt. 00:06:58.779 --> 00:07:01.180 Okay. And the sources mention this needs to be 00:07:01.180 --> 00:07:03.160 really structured, not just a paragraph of text. 00:07:03.659 --> 00:07:07.220 Absolutely. That structure is like your defense 00:07:07.220 --> 00:07:11.120 against prompt drift, where the AI starts forgetting 00:07:11.120 --> 00:07:13.480 its instructions over time or gets sidetracked. 00:07:13.519 --> 00:07:15.160 Makes sense. So what's the structure look like? 00:07:15.339 --> 00:07:17.939 It starts with an overview, basically telling 00:07:17.939 --> 00:07:20.540 the AI who it is, like you are an expert technical 00:07:20.540 --> 00:07:23.680 stock analyst. Then comes context listing the 00:07:23.680 --> 00:07:26.250 tools it has access to. like our chart specialist, 00:07:26.470 --> 00:07:29.250 and maybe what data sources it can use. After 00:07:29.250 --> 00:07:31.649 that, step -by -step instructions on how to handle 00:07:31.649 --> 00:07:34.189 requests. Okay, overview, context, instructions, 00:07:34.410 --> 00:07:37.709 what else? Then, importantly, detailed tool descriptions. 00:07:38.069 --> 00:07:40.769 So, explicitly naming the get chart tool and 00:07:40.769 --> 00:07:43.730 maybe explaining what it does. And finally, SOPs 00:07:43.730 --> 00:07:46.139 as standard operating procedures. This is where 00:07:46.139 --> 00:07:48.220 you repeat the absolute most critical rules, 00:07:48.339 --> 00:07:50.899 like the big one. Do not give explicit financial 00:07:50.899 --> 00:07:53.120 advice. You hammer that home. That structured 00:07:53.120 --> 00:07:55.879 approach seems vital. Oh, it is, honestly. I 00:07:55.879 --> 00:07:57.839 still wrestle with prompt drift myself sometimes. 00:07:58.100 --> 00:07:59.720 Keeping things tightly structured like this, 00:07:59.759 --> 00:08:01.899 especially with those SOPs, is pretty essential 00:08:01.899 --> 00:08:04.180 for keeping the agent reliable and focused query 00:08:04.180 --> 00:08:07.339 after query. So let's connect that back. If we're 00:08:07.339 --> 00:08:10.459 building this scalable no -code system, Why is 00:08:10.459 --> 00:08:11.800 going through the trouble of using something 00:08:11.800 --> 00:08:14.480 like the Alpha Arena to pick maybe a less common 00:08:14.480 --> 00:08:18.240 model like DeepSeek V3 .1 so important? Why not 00:08:18.240 --> 00:08:21.560 just default to, say, GPT -4 or Claude, which 00:08:21.560 --> 00:08:23.839 everyone knows? Yeah, good question. It comes 00:08:23.839 --> 00:08:25.660 back to that performance under pressure idea. 00:08:26.040 --> 00:08:28.680 The Alpha Arena helps select the model that's 00:08:28.680 --> 00:08:31.699 proven to be most reliable and accurate specifically 00:08:31.699 --> 00:08:35.340 for this kind of dynamic reasoning task. In finance, 00:08:35.440 --> 00:08:37.220 you know, that reliability, that accuracy when 00:08:37.220 --> 00:08:39.379 things get volatile, that trumps familiarity 00:08:39.379 --> 00:08:41.659 every single time. You want the best cool for 00:08:41.659 --> 00:08:43.960 that specific brain function. OK, so the CEO 00:08:43.960 --> 00:08:46.019 brain running on DeepSeek decides it needs a 00:08:46.019 --> 00:08:48.500 chart and some analysis for, say, Apple. How 00:08:48.500 --> 00:08:50.840 does it actually hand off that specific task 00:08:50.840 --> 00:08:53.679 to the specialist tool, that sub workflow? Right. 00:08:53.740 --> 00:08:55.259 So the brain basically calls the specialist. 00:08:55.370 --> 00:08:58.850 tool using its unique url that call triggers 00:08:58.850 --> 00:09:01.429 the sub workflow to start running like dialing 00:09:01.429 --> 00:09:04.190 its direct number kind of yeah and the absolute 00:09:04.190 --> 00:09:07.730 key detail here for setup is the naming the name 00:09:07.730 --> 00:09:10.169 you give the sub workflow in your no code tool 00:09:10.649 --> 00:09:13.210 Let's say you call it get chart. That name has 00:09:13.210 --> 00:09:16.330 to exactly match the tool name you defined back 00:09:16.330 --> 00:09:18.730 in the brain system prompt. Exactly match, like 00:09:18.730 --> 00:09:20.850 capitalization and everything. Everything. Case 00:09:20.850 --> 00:09:23.549 sensitive, spaces, underscores, whatever. If 00:09:23.549 --> 00:09:26.309 it's off by even one character, the brain basically 00:09:26.309 --> 00:09:28.509 says, I don't know what tool you're talking about. 00:09:28.610 --> 00:09:30.610 And the whole process breaks down. Connection 00:09:30.610 --> 00:09:33.110 failed. Okay, meticulous naming is critical. 00:09:33.409 --> 00:09:36.090 So the specialist gets the call. Its first job 00:09:36.090 --> 00:09:38.250 is getting the actual chart image, right? It 00:09:38.250 --> 00:09:42.009 uses that free service chart IMG ATI. Yep, that's 00:09:42.009 --> 00:09:44.049 the one. And the sources pointed out a really 00:09:44.049 --> 00:09:46.049 neat efficiency trick here for setting up that 00:09:46.049 --> 00:09:48.330 API call in NANN. Ah, yeah, they called it the 00:09:48.330 --> 00:09:51.789 pro gamer move using the CRL import method. Uh 00:09:51.789 --> 00:09:54.450 -huh, yeah, that's the one. And it's genuinely 00:09:54.450 --> 00:09:56.710 a massive time saver. So what does that trick 00:09:56.710 --> 00:09:59.450 actually do for the person? Building this. How 00:09:59.450 --> 00:10:01.769 does it simplify things? Well, instead of manually 00:10:01.769 --> 00:10:03.929 setting up every single parameter for the API 00:10:03.929 --> 00:10:06.649 call, you know, the headers, the request body, 00:10:06.789 --> 00:10:09.149 the authentication method, which can be super 00:10:09.149 --> 00:10:11.309 tedious and error prone. Right. Fiddling with 00:10:11.309 --> 00:10:13.450 JSON and stuff. Exactly. You just find the example 00:10:13.450 --> 00:10:16.389 CRL command in the chart IMG API documentation. 00:10:16.769 --> 00:10:19.250 It's usually just a snippet of text. You copy 00:10:19.250 --> 00:10:21.570 it and you paste it into this specific import 00:10:21.570 --> 00:10:24.850 function in the NNN HTTP request node. And then? 00:10:25.450 --> 00:10:27.830 Bam. And it automatically configures the entire 00:10:27.830 --> 00:10:29.710 node for you. So that's all the headers, parameters, 00:10:29.889 --> 00:10:32.750 everything based on that URL command. It literally 00:10:32.750 --> 00:10:35.889 turns what could be hours of debugging into like 00:10:35.889 --> 00:10:39.309 10 seconds of copy paste. Wow. Okay. That is 00:10:39.309 --> 00:10:42.409 a huge shortcut. So when configuring that API 00:10:42.409 --> 00:10:45.250 request to chart IMG, what are the key settings? 00:10:45.509 --> 00:10:47.909 Two main things mentioned. First, using storage 00:10:47.909 --> 00:10:50.929 in the API's URL path. That tells Chart .img 00:10:50.929 --> 00:10:53.110 to generate the chart and give you back a public 00:10:53.110 --> 00:10:55.389 URL link to the image file, which is what we 00:10:55.389 --> 00:10:57.210 need. Okay, so you get a link to the picture. 00:10:57.429 --> 00:10:59.690 Yep. And second, you need to dynamically insert 00:10:59.690 --> 00:11:01.889 the stock ticker the user actually asked for. 00:11:02.090 --> 00:11:04.570 Use an N8N expression for that, something like 00:11:04.570 --> 00:11:07.230 trainison .query. That pulls the ticker, say 00:11:07.230 --> 00:11:10.210 NASDAQ key geog, from the data the brain sent 00:11:10.210 --> 00:11:13.129 over and sticks it into the API call. Got it. 00:11:13.169 --> 00:11:15.629 So the chart gets generated. We get a URL back. 00:11:16.110 --> 00:11:18.929 What's step three? Step three is actually downloading 00:11:18.929 --> 00:11:21.909 that image. You need the image file, not just 00:11:21.909 --> 00:11:24.730 the link, to feed to the Vision AI later. So 00:11:24.730 --> 00:11:27.190 that requires a second HTTP request node. Another 00:11:27.190 --> 00:11:29.690 API call. Yeah, this one just takes the URL from 00:11:29.690 --> 00:11:33.029 step two and fetches the actual image data. And 00:11:33.029 --> 00:11:35.009 the crucial setting here is you have to set the 00:11:35.009 --> 00:11:38.179 response format to binary. Binary. Okay, what 00:11:38.179 --> 00:11:40.399 exactly is binary data here? Think of binary 00:11:40.399 --> 00:11:43.240 data as the raw file format for the image itself, 00:11:43.360 --> 00:11:45.820 the actual pixels and colors that an AI vision 00:11:45.820 --> 00:11:48.279 model can understand. Not text. Right. If you 00:11:48.279 --> 00:11:50.299 forget to set it to binary, the node just downloads 00:11:50.299 --> 00:11:52.679 the image link as text, and when you pass that 00:11:52.679 --> 00:11:55.259 text to the vision AI, well, it can't see text. 00:11:55.340 --> 00:11:57.759 It needs the picture. Analysis fails instantly. 00:11:58.250 --> 00:12:01.269 Okay, binary format is key for the image download. 00:12:01.549 --> 00:12:04.490 And that brings us to step four, the actual specialized 00:12:04.490 --> 00:12:08.470 analysis using AI vision. And this is where that 00:12:08.470 --> 00:12:10.629 dual AI strategy really comes into play, right? 00:12:10.710 --> 00:12:13.169 Exactly. Because even though our brain is deep 00:12:13.169 --> 00:12:15.649 seek, for this specific task, analyzing the visual 00:12:15.649 --> 00:12:19.289 chart, we switch models, we use GPT -4O. Ah, 00:12:19.350 --> 00:12:22.149 interesting. Why the switch? Because, well... 00:12:22.350 --> 00:12:24.269 Currently, GPT -4 is generally considered the 00:12:24.269 --> 00:12:26.529 best model out there, specifically for visual 00:12:26.529 --> 00:12:29.029 analysis, for understanding images. So you're 00:12:29.029 --> 00:12:32.029 using the best tool for each distinct job, DeepSeq 00:12:32.029 --> 00:12:34.809 for reasoning and managing, GPT -4 for seeing. 00:12:35.289 --> 00:12:37.450 Precisely. We're essentially paying a bit more 00:12:37.450 --> 00:12:39.429 maybe for that specific vision step, but we're 00:12:39.429 --> 00:12:41.470 outsourcing that really critical high -value 00:12:41.470 --> 00:12:43.669 task, interpreting the candlestick patterns, 00:12:43.830 --> 00:12:46.830 the Mali lines, the volume bars visually to the 00:12:46.830 --> 00:12:49.029 absolute best specialist model available for 00:12:49.029 --> 00:12:51.100 that job. That makes a lot of sense. Pick the 00:12:51.100 --> 00:12:53.860 expert for the specific task. It really demonstrates 00:12:53.860 --> 00:12:58.139 the power of this modular approach. Whoa. Just 00:12:58.139 --> 00:13:00.600 imagine scaling this kind of architecture, like 00:13:00.600 --> 00:13:04.500 across a huge company, using the absolute best 00:13:04.500 --> 00:13:07.759 AI model for every single tiny step in a really 00:13:07.759 --> 00:13:10.370 complex business workflow. That feels like the 00:13:10.370 --> 00:13:12.009 future of building AI agents, doesn't it? It 00:13:12.009 --> 00:13:14.049 really does. So when you feed that downloaded 00:13:14.049 --> 00:13:16.909 chart image to GPT -40, what do you ask it to 00:13:16.909 --> 00:13:18.889 look for? The prompt must be pretty specific. 00:13:19.029 --> 00:13:20.649 Oh, yeah. Super specific. You don't just say 00:13:20.649 --> 00:13:22.950 analyze this chart. You instruct it to perform 00:13:22.950 --> 00:13:26.409 distinct types of analysis. Yeah. Detailed candlestick 00:13:26.409 --> 00:13:29.490 pattern analysis, MCG analysis, like is it crossing 00:13:29.490 --> 00:13:32.429 over? Volume analysis, are there spikes? And 00:13:32.429 --> 00:13:35.309 crucially, to identify and map the key support 00:13:35.309 --> 00:13:37.610 and resistance levels directly onto what's understanding 00:13:37.610 --> 00:13:39.710 the chart. So you're forcing it to act like a 00:13:39.710 --> 00:13:41.629 real technical analyst, looking for those specific 00:13:41.629 --> 00:13:43.529 indicators, not just describing the picture. 00:13:43.710 --> 00:13:46.110 Exactly. You want a professional readout, actionable 00:13:46.110 --> 00:13:48.960 insights derived from the visual data. So drilling 00:13:48.960 --> 00:13:50.840 down on that, what is the biggest advantage then 00:13:50.840 --> 00:13:53.320 of using an AI vision model to look at the chart 00:13:53.320 --> 00:13:56.480 image compared to just feeding a regular text 00:13:56.480 --> 00:13:59.840 -based AI model the raw price and volume data 00:13:59.840 --> 00:14:02.620 in tables? Great question. It's about pattern 00:14:02.620 --> 00:14:05.000 recognition, really. The vision model can instantly 00:14:05.000 --> 00:14:08.700 see complex visual patterns like a bullish engulfing 00:14:08.700 --> 00:14:11.240 candlestick pattern or the specific shape of 00:14:11.240 --> 00:14:13.759 a Mackie divergence or a sudden volume spike 00:14:13.759 --> 00:14:16.299 coinciding with a price move. It can recognize 00:14:16.299 --> 00:14:18.899 these holistic visual signals. much faster and 00:14:18.899 --> 00:14:21.700 often more reliably than a text model trying 00:14:21.700 --> 00:14:24.200 to infer those same patterns just from raw numbers 00:14:24.200 --> 00:14:26.460 in a table. It literally sees the picture the 00:14:26.460 --> 00:14:28.820 way a human analyst would. So we put it all together. 00:14:29.389 --> 00:14:31.730 User sends a message on Telegram. The brain, 00:14:31.889 --> 00:14:33.529 deep -seek, figures out what's needed, maybe 00:14:33.529 --> 00:14:35.649 calls a specialist sub -workflow. The specialist, 00:14:35.850 --> 00:14:39.070 using chart IMG API, generates the chart image 00:14:39.070 --> 00:14:42.289 URL, then downloads the binary image data. Right. 00:14:42.350 --> 00:14:45.129 Then that image is sent to the vision AI, GPT 00:14:45.129 --> 00:14:48.250 -4, with specific analysis instructions. GPT 00:14:48.250 --> 00:14:50.990 -4 sends back the detailed text analysis. Which 00:14:50.990 --> 00:14:53.110 goes back to the specialist, then back up to 00:14:53.110 --> 00:14:56.389 the brain. The brain takes that analysis, maybe 00:14:56.389 --> 00:14:58.649 combines it with conversation history or other 00:14:58.649 --> 00:15:01.149 analysis if it was a comparison request. And 00:15:01.149 --> 00:15:03.669 finally formats it all into a nice message and 00:15:03.669 --> 00:15:06.149 sends it back to the user on Telegram. Exactly. 00:15:06.149 --> 00:15:08.950 And the amazing thing is that whole round trip. 00:15:09.500 --> 00:15:11.639 Apparently it takes only about 10 to 20 seconds. 00:15:11.879 --> 00:15:14.700 Wow, that's remarkably fast for that much processing 00:15:14.700 --> 00:15:17.259 and multiple AI calls. It really is. And that 00:15:17.259 --> 00:15:19.700 architecture really proves its worth with that 00:15:19.700 --> 00:15:21.700 multi -stock comparison feature we talked about 00:15:21.700 --> 00:15:23.840 earlier. Right, comparing Alphabet and Microsoft. 00:15:24.240 --> 00:15:26.120 Yeah. The brain is smart enough to recognize 00:15:26.120 --> 00:15:29.440 this needs two separate analyses. So it calls 00:15:29.440 --> 00:15:32.580 a socialist tool twice, once for each stock. 00:15:32.860 --> 00:15:35.159 Gets back two sets of chart analysis texts from 00:15:35.159 --> 00:15:39.559 GPT -4. And then... The brain performs that high 00:15:39.559 --> 00:15:42.779 -level reasoning itself using DeepSeq to synthesize 00:15:42.779 --> 00:15:44.919 those two reports into that final comparative 00:15:44.919 --> 00:15:47.639 summary for the user. It's a real orchestration 00:15:47.639 --> 00:15:49.940 feat, managing that flow. Definitely sounds like 00:15:49.940 --> 00:15:52.360 it. Now, are there any limitations or things 00:15:52.360 --> 00:15:55.179 to watch out for with this specific build? Yeah, 00:15:55.240 --> 00:15:57.480 a couple of key ones mentioned. First, that free 00:15:57.480 --> 00:16:00.740 chart IMG API they used. It's limited. It only 00:16:00.740 --> 00:16:03.340 works for NASDAQ -listed securities. Ah, okay. 00:16:03.440 --> 00:16:06.440 So if you ask for a stock on the London Stock 00:16:06.440 --> 00:16:08.740 Exchange or something. The sub -workflow will 00:16:08.740 --> 00:16:11.620 probably just fail, hopefully gracefully, but 00:16:11.620 --> 00:16:14.259 you won't get a chart or analysis back. That's 00:16:14.259 --> 00:16:17.399 a limitation of that specific free tool. Good 00:16:17.399 --> 00:16:19.700 to know. What about common mistakes people make 00:16:19.700 --> 00:16:21.559 when trying to build something like this? Two 00:16:21.559 --> 00:16:23.950 big ones were highlighted. And they're easy traps 00:16:23.950 --> 00:16:26.970 to fall into. First, we mentioned it, but it's 00:16:26.970 --> 00:16:30.509 worth repeating. The tool name. The name you 00:16:30.509 --> 00:16:32.850 give your specialist workflow, like GetChart 00:16:32.850 --> 00:16:35.769 in your no -code platform, must perfectly, exactly 00:16:35.769 --> 00:16:38.110 match the tool name you wrote in the brain system 00:16:38.110 --> 00:16:40.669 prompt. Case -sensitive spaces, the whole deal. 00:16:40.850 --> 00:16:43.110 Got it. Get it wrong, and the brain can't call 00:16:43.110 --> 00:16:45.559 the specialist dead end. Nah. The second big 00:16:45.559 --> 00:16:47.740 one, and we also touched on this, is ensuring 00:16:47.740 --> 00:16:50.860 the chat ID variable. From Telegram. Yes, making 00:16:50.860 --> 00:16:52.960 absolute sure it's correctly mapped not just 00:16:52.960 --> 00:16:54.980 to the memory node, but also to all the Telegram 00:16:54.980 --> 00:16:57.740 send nodes that reply back to the user. Why all 00:16:57.740 --> 00:17:00.059 the send nodes too? Because if you only link 00:17:00.059 --> 00:17:03.179 it to memory, but not the reply node, the agent 00:17:03.179 --> 00:17:05.279 might remember the right conversation, but it 00:17:05.279 --> 00:17:07.380 could still accidentally send the reply to the 00:17:07.380 --> 00:17:10.299 wrong user's chat window. You need that chat 00:17:10.299 --> 00:17:13.160 ID guiding both memory and replies to keep everything 00:17:13.160 --> 00:17:16.019 strictly tied to the right conversation. Missing 00:17:16.019 --> 00:17:17.900 that is how you get those frustrating crossed 00:17:17.900 --> 00:17:21.220 messages. Okay, so map the chat ID everywhere 00:17:21.220 --> 00:17:24.299 it relates to a specific user interaction, memory, 00:17:24.519 --> 00:17:27.359 and sending messages back. Bingo. Nail those 00:17:27.359 --> 00:17:29.579 two things, exact tool names and consistent chat 00:17:29.579 --> 00:17:33.740 ID mapping, and you avoid probably... 80 % of 00:17:33.740 --> 00:17:35.579 the common frustrations. And all this efficiency, 00:17:35.839 --> 00:17:38.420 this complex orchestration, it really brings 00:17:38.420 --> 00:17:40.660 us back to the power of no -code advantage, doesn't 00:17:40.660 --> 00:17:42.839 it? Absolutely. I mean, think about it. This 00:17:42.839 --> 00:17:45.599 system connects multiple different APIs, manages 00:17:45.599 --> 00:17:48.119 conversational state with memory, strategically 00:17:48.119 --> 00:17:50.559 swaps between different best -in -class AI models 00:17:50.559 --> 00:17:53.000 for different tasks. And delivers results in 00:17:53.000 --> 00:17:55.880 seconds. Yeah. And it was all built. Visually, 00:17:55.900 --> 00:17:58.960 in maybe a few hours, on a platform like NAN, 00:17:58.960 --> 00:18:01.559 a traditional software developer trying to code 00:18:01.559 --> 00:18:03.920 this from scratch, setting up the servers, the 00:18:03.920 --> 00:18:06.420 API integrations, the state management, the model 00:18:06.420 --> 00:18:09.180 switching logic, that could easily be a multi 00:18:09.180 --> 00:18:11.539 -week, maybe even multi -month project. It really 00:18:11.539 --> 00:18:13.500 lowers the barrier to creating sophisticated 00:18:13.500 --> 00:18:16.619 AI tools. Fundamentally. It democratizes the 00:18:16.619 --> 00:18:19.519 ability to build these complex, specialized AI 00:18:19.519 --> 00:18:22.069 agents. That's the real power here. Hashtag, 00:18:22.069 --> 00:18:24.069 hashtag, hashtag outro. So reflecting on this 00:18:24.069 --> 00:18:25.930 whole project, it feels like it demonstrates 00:18:25.930 --> 00:18:28.650 three really critical ideas for building advanced 00:18:28.650 --> 00:18:32.250 AI systems now. First, that separation of concerns. 00:18:32.690 --> 00:18:34.950 Using that brain specialist architecture, the 00:18:34.950 --> 00:18:37.250 CEO and the expert consultant model, it just 00:18:37.250 --> 00:18:39.670 makes systems cleaner, more modular and much 00:18:39.670 --> 00:18:42.109 easier to scale later on. Yeah, definitely. Second 00:18:42.109 --> 00:18:44.660 big takeaway, AI vision is a game changer. The 00:18:44.660 --> 00:18:47.039 ability for AI to actually analyze images like 00:18:47.039 --> 00:18:49.460 these stock charts unlocks totally new possibilities. 00:18:49.660 --> 00:18:51.940 It lets the agent see the market in a way text 00:18:51.940 --> 00:18:55.000 alone can't capture. It moves beyond just processing 00:18:55.000 --> 00:18:57.500 numbers. And third, it highlights that tools 00:18:57.500 --> 00:19:00.460 make the agent. An AI agent isn't just defined 00:19:00.460 --> 00:19:03.039 by its conversational ability anymore. It's defined 00:19:03.039 --> 00:19:05.839 by the specialized tools it can connect to and 00:19:05.839 --> 00:19:08.480 use effectively to perform complex multi -step 00:19:08.480 --> 00:19:10.720 actions out in the world. That's so true. And, 00:19:10.779 --> 00:19:12.799 you know, this specific brain plus specialist. 00:19:13.740 --> 00:19:15.759 tool architecture. It's not just for stocks. 00:19:15.940 --> 00:19:18.799 You can adapt this exact model for countless 00:19:18.799 --> 00:19:21.410 other specialized uses. Like what? Oh, I don't 00:19:21.410 --> 00:19:23.630 know. Maybe a financial advisor tool that pulls 00:19:23.630 --> 00:19:26.410 specific compliance documents or a trading education 00:19:26.410 --> 00:19:28.490 bot that can analyze user submitted practice 00:19:28.490 --> 00:19:31.849 trades or even hyper -focused content creation 00:19:31.849 --> 00:19:34.789 tools for any niche industry where you have specialized 00:19:34.789 --> 00:19:37.849 data sources or analysis steps. The pattern is 00:19:37.849 --> 00:19:40.369 reusable. That's a really powerful idea. So maybe 00:19:40.369 --> 00:19:42.390 the final thought for everyone listening is this. 00:19:42.509 --> 00:19:45.250 If you can build an expert technical stock analyst 00:19:45.250 --> 00:19:48.259 with no code. What other specialized tools, unique 00:19:48.259 --> 00:19:51.859 data sources, or expert processes exist in your 00:19:51.859 --> 00:19:54.079 field? Yeah, what could you connect an AI agent 00:19:54.079 --> 00:19:56.500 brain to? Because the future here seems to be 00:19:56.500 --> 00:19:58.960 less about just building a better chatbot and 00:19:58.960 --> 00:20:01.480 more about building better connections, orchestrating 00:20:01.480 --> 00:20:04.359 these specialized tools to achieve complex, valuable 00:20:04.359 --> 00:20:07.599 tasks. It's about connection and action, not 00:20:07.599 --> 00:20:08.259 just conversation.