WEBVTT 00:00:00.000 --> 00:00:03.359 Welcome to the Deep Dive. Today, we're plunging 00:00:03.359 --> 00:00:05.980 into a topic that's, well, it's not science fiction 00:00:05.980 --> 00:00:08.140 anymore, is it? Artificial intelligence. Not 00:00:08.140 --> 00:00:12.220 at all. It's woven into everything we do. Recommendations, 00:00:12.419 --> 00:00:15.939 voice assistance. It's just there. It really 00:00:15.939 --> 00:00:18.140 is. But you know, with AI booming, there's this 00:00:18.140 --> 00:00:21.420 just ocean of information out there. Courses, 00:00:21.480 --> 00:00:23.940 videos, tutorials. Uh -huh. It's overwhelming. 00:00:24.219 --> 00:00:26.949 Yeah. And it's weird because all that information 00:00:26.949 --> 00:00:29.469 can actually push people into what some call 00:00:29.469 --> 00:00:32.049 tutorial hell. Well, I know it well. You watch 00:00:32.049 --> 00:00:34.170 hours of stuff, learn bits and pieces. Right. 00:00:34.390 --> 00:00:37.950 Isolated concepts, but you never quite connect 00:00:37.950 --> 00:00:39.829 the dots. You don't build a solid skill set. 00:00:39.950 --> 00:00:42.329 It's like having puzzle pieces scattered everywhere, 00:00:42.490 --> 00:00:44.890 but no final picture. Exactly. You feel busy, 00:00:44.929 --> 00:00:46.770 but you're not really building anything substantial. 00:00:47.250 --> 00:00:49.450 So that's our mission today. We want to give 00:00:49.450 --> 00:00:52.049 you a kind of strategic roadmap, not just a list 00:00:52.049 --> 00:00:54.009 of things to learn, but a way to think about 00:00:54.009 --> 00:00:55.890 it, building from the ground up. Yeah, moving 00:00:55.890 --> 00:00:58.909 from just being an AI user to becoming an AI 00:00:58.909 --> 00:01:01.649 builder, understanding the why behind each step. 00:01:02.090 --> 00:01:04.790 Precisely. Focusing on the principles. So let's 00:01:04.790 --> 00:01:07.329 dive in. Sounds good. That tutorial hell you 00:01:07.329 --> 00:01:09.170 mentioned, it's been there so many hours, felt 00:01:09.170 --> 00:01:11.750 like I was learning, but yeah, like you said, 00:01:12.150 --> 00:01:13.829 disconnected pieces. What's the biggest trap 00:01:13.829 --> 00:01:18.290 you see people fall into? Probably jumping straight 00:01:18.290 --> 00:01:21.510 to the really complex flashy AI models without 00:01:21.510 --> 00:01:24.810 getting the basics down first. It's like trying 00:01:24.810 --> 00:01:27.129 to build a skyscraper without pouring the foundation. 00:01:27.329 --> 00:01:29.530 It just won't stand. That totally nails it. So 00:01:29.530 --> 00:01:31.870 if you really want to build with AI, not just 00:01:31.870 --> 00:01:34.629 use it, the first thing, the absolute non -negotiable 00:01:34.629 --> 00:01:37.890 bedrock. It has to be solid programming and software 00:01:37.890 --> 00:01:40.069 engineering skills. Definitely. Because an AI 00:01:40.069 --> 00:01:43.060 model, no matter how cool, it's still just one 00:01:43.060 --> 00:01:45.019 part of a bigger software system. Right. And 00:01:45.019 --> 00:01:47.200 if you don't have those core software skills, 00:01:47.640 --> 00:01:50.239 your amazing model is basically, well, trapped 00:01:50.239 --> 00:01:52.780 in its Jupiter notebook. Stuck. Can't be deployed, 00:01:52.980 --> 00:01:55.099 can't be integrated, can't solve real problems 00:01:55.099 --> 00:01:57.400 out in the world. So why is that software foundation 00:01:57.400 --> 00:01:59.760 so critical? What does it actually do for you? 00:01:59.939 --> 00:02:03.540 OK, three big reasons, I think. First, maintainability. 00:02:04.159 --> 00:02:06.920 AI code isn't a one -off thing. You're always 00:02:06.920 --> 00:02:09.800 improving it, debugging. It's alive, almost. 00:02:10.159 --> 00:02:12.949 Yeah. And good engineering practices, clean code, 00:02:13.569 --> 00:02:15.629 that stops it becoming a total nightmare to manage 00:02:15.629 --> 00:02:18.930 later. OK, maintainability. What else? Scalability. 00:02:19.469 --> 00:02:21.250 You got to design systems that can handle real 00:02:21.250 --> 00:02:25.030 demand, like maybe 1 ,000 requests per second 00:02:25.030 --> 00:02:28.409 hitting your model. Wow. OK. That's serious scale. 00:02:28.569 --> 00:02:30.870 It is. And software engineering teaches you how 00:02:30.870 --> 00:02:32.330 to build for that. Otherwise, it just buckles 00:02:32.330 --> 00:02:36.509 under pressure. And third, collaboration. Ah, 00:02:36.729 --> 00:02:39.009 working in teams. Exactly. You're almost never 00:02:39.009 --> 00:02:41.370 building AI alone on serious projects. So you 00:02:41.370 --> 00:02:44.030 need standardized code, shared practices, tools 00:02:44.030 --> 00:02:46.650 like Git. It's just chaos without it. That makes 00:02:46.650 --> 00:02:48.909 sense. You mentioned maintainability and scalability. 00:02:49.229 --> 00:02:51.250 Are there common mistakes people make there, 00:02:51.349 --> 00:02:53.210 things that look okay at first but cause problems? 00:02:53.229 --> 00:02:56.889 Oh, for sure. A big one is underestimating Python's 00:02:56.889 --> 00:02:59.189 ecosystem. People learn Python the language, 00:02:59.229 --> 00:03:01.530 but they don't always grasp the power of its 00:03:01.530 --> 00:03:04.460 libraries. Like NumPy and Pandas. Exactly. It's 00:03:04.460 --> 00:03:07.439 the batteries included thing. Take NumPy. It's 00:03:07.439 --> 00:03:09.900 not just a ray tricks. It's the vectorized computation. 00:03:10.219 --> 00:03:12.539 That's the engine under the hood for performance 00:03:12.539 --> 00:03:15.500 in deep learning. When you're processing millions 00:03:15.500 --> 00:03:18.060 of numbers. Efficiency is everything. It's critical, 00:03:18.099 --> 00:03:21.099 yeah. And pandas. Almost every ML project dealing 00:03:21.099 --> 00:03:24.340 with spreadsheets or tables starts with pandas. 00:03:24.560 --> 00:03:28.490 Data wrangling, cleaning. It's essential. OK, 00:03:28.629 --> 00:03:30.710 so mastering the ecosystem, not just the language 00:03:30.710 --> 00:03:33.629 itself, what about core computer science stuff 00:03:33.629 --> 00:03:36.610 like data structures and algorithms? People sometimes 00:03:36.610 --> 00:03:39.669 think that's just for job interviews. Huge mistake 00:03:39.669 --> 00:03:42.569 to skip it, honestly. DSNA isn't just theory. 00:03:42.930 --> 00:03:45.509 It teaches you how to write efficient code. How 00:03:45.509 --> 00:03:47.849 does that connect directly to AI building? Well, 00:03:48.009 --> 00:03:51.110 simple example. Knowing when to use a dictionary, 00:03:51.310 --> 00:03:53.310 a hash map, instead of just a list for looking 00:03:53.310 --> 00:03:55.729 stuff up, that choice can speed up your data 00:03:55.729 --> 00:03:58.650 pre -processing step by, like, thousands of times. 00:03:58.830 --> 00:04:00.870 Seriously. Thousands. Yeah, it's not academic. 00:04:01.069 --> 00:04:02.550 That's a real difference I can make or break 00:04:02.550 --> 00:04:04.610 whether your AI is actually usable in the real 00:04:04.610 --> 00:04:08.689 world. Wow. Okay, so efficiency matters. And 00:04:08.689 --> 00:04:10.949 you mentioned collaboration earlier, which brings 00:04:10.949 --> 00:04:14.770 us to Git, right? Version control. Oh, absolutely 00:04:14.770 --> 00:04:17.470 non -negotiable. It's just a requirement, full 00:04:17.470 --> 00:04:20.720 stop. Especially in AI, reproducibility is key. 00:04:20.879 --> 00:04:23.420 Because you're always experimenting. Constantly. 00:04:23.920 --> 00:04:26.500 Git lets you track every single change code, 00:04:26.759 --> 00:04:29.300 data, parameters. You can easily jump back to 00:04:29.300 --> 00:04:32.120 any version you tried. Plus, yeah, for teams, 00:04:32.540 --> 00:04:34.180 it's how you coordinate without driving each 00:04:34.180 --> 00:04:36.709 other crazy. Okay, so... Programming and software 00:04:36.709 --> 00:04:39.589 engineering is the bedrock, the tool. What's 00:04:39.589 --> 00:04:41.569 next? Well, if programming is the tool, then 00:04:41.569 --> 00:04:43.550 math is kind of the language you use with that 00:04:43.550 --> 00:04:45.910 tool. It's how you talk to the data and the models. 00:04:46.269 --> 00:04:48.810 Right. Understanding the math moves you beyond 00:04:48.810 --> 00:04:51.129 just typing model .fit and hoping for the best. 00:04:51.370 --> 00:04:53.410 Exactly. It lets you actually understand research 00:04:53.410 --> 00:04:56.829 papers, grasp why a model works or doesn't work, 00:04:56.870 --> 00:04:59.050 and figure out how to fix it when things go wrong, 00:04:59.370 --> 00:05:01.910 which they inevitably do. So where do we start 00:05:01.910 --> 00:05:04.750 with the math? Linear algebra. That's fundamental. 00:05:04.949 --> 00:05:08.720 Think about it. All AI, data, images, text, sound, 00:05:09.019 --> 00:05:10.860 it all gets turned into numbers. Right, into 00:05:10.860 --> 00:05:14.319 vectors and matrices. Precisely. Linear algebra 00:05:14.319 --> 00:05:16.399 gives you the language and the tools to describe 00:05:16.399 --> 00:05:19.420 those structures and, crucially, to manipulate 00:05:19.420 --> 00:05:22.160 them. It's how you tell the model what it's looking 00:05:22.160 --> 00:05:24.620 at and how to transform it. OK, so linear algebra 00:05:24.620 --> 00:05:27.019 describes the data structure. What about the 00:05:27.019 --> 00:05:28.759 learning part? How do models actually learn? 00:05:29.000 --> 00:05:32.230 That's where calculus comes in. Calculus, specifically 00:05:32.230 --> 00:05:35.189 differentiation, is the engine of learning for 00:05:35.189 --> 00:05:37.470 most models. You mean gradient descent? Yep. 00:05:37.889 --> 00:05:40.589 That core idea of finding the minimum error by 00:05:40.589 --> 00:05:43.550 following the slope, the derivative. That's calculus. 00:05:44.370 --> 00:05:46.850 Understanding how learning rates work or different 00:05:46.850 --> 00:05:49.269 optimization algorithms like Atom, it all comes 00:05:49.269 --> 00:05:52.310 down to navigating that error landscape using 00:05:52.310 --> 00:05:54.610 calculus. So it's not just knowing it minimizes 00:05:54.610 --> 00:05:57.519 error, but how it does it. Exactly. that lets 00:05:57.519 --> 00:06:00.019 you tune it properly. Okay, linear algebra for 00:06:00.019 --> 00:06:02.480 structure, calculus for learning. What else? 00:06:02.819 --> 00:06:04.879 Statistics and probability. This is your toolkit 00:06:04.879 --> 00:06:07.139 for the real world because the real world is 00:06:07.139 --> 00:06:09.759 messy. It's full of randomness and uncertainty. 00:06:09.860 --> 00:06:12.259 How does that apply day to day? Well, evaluating 00:06:12.259 --> 00:06:15.040 your models properly. Knowing when simple accuracy 00:06:15.040 --> 00:06:16.980 isn't good enough, right? Like with unbalanced 00:06:16.980 --> 00:06:20.379 data, you need metrics like F1 score or AUC. 00:06:20.519 --> 00:06:22.959 Statistics tells you why and when. Okay, evaluation. 00:06:23.259 --> 00:06:26.379 What else? Understanding your data itself, distributions, 00:06:26.800 --> 00:06:29.560 outliers, weird patterns, statistics helps you 00:06:29.560 --> 00:06:32.560 make sense of it. And really importantly, A -B 00:06:32.560 --> 00:06:35.180 testing. To prove your new model is actually 00:06:35.180 --> 00:06:38.459 better. Yes, scientifically prove it, not just 00:06:38.459 --> 00:06:41.000 guess. Statistics provides the framework for 00:06:41.000 --> 00:06:43.660 that. OK, so we've got the bedrock, programming 00:06:43.660 --> 00:06:45.879 and software engineering. We've got the language, 00:06:46.399 --> 00:06:48.959 math and stats. Now we can actually start building 00:06:48.959 --> 00:06:50.860 the intelligent systems, right? Now we get to 00:06:50.860 --> 00:06:53.500 the main pillar. machine learning itself. And 00:06:53.500 --> 00:06:55.839 you mentioned earlier, common mistake is jumping 00:06:55.839 --> 00:06:58.279 straight past this to deep learning. Yeah, it 00:06:58.279 --> 00:07:01.980 happens a lot. But honestly, a huge chunk of 00:07:01.980 --> 00:07:04.279 real world business problems think forecasting 00:07:04.279 --> 00:07:07.019 sales, predicting if a customer will leave, stuff 00:07:07.019 --> 00:07:09.600 usually based on tabular data like spreadsheets. 00:07:09.740 --> 00:07:11.600 Classic machine learning models like gradient 00:07:11.600 --> 00:07:13.860 boosting are often the best tool for the job. 00:07:14.180 --> 00:07:16.839 They can be faster, need less data, and they're 00:07:16.839 --> 00:07:18.819 often much easier to understand and explain. 00:07:19.579 --> 00:07:22.720 So it's not always about using the most complex 00:07:22.720 --> 00:07:25.319 tool available. Definitely not. But even with 00:07:25.319 --> 00:07:27.480 a classic ML, it's way more than just calling 00:07:27.480 --> 00:07:30.939 that model .fit function we mentioned. Ah, right. 00:07:31.000 --> 00:07:33.300 The whole lifecycle. Exactly. There's a standard 00:07:33.300 --> 00:07:35.579 process. It usually starts with getting the data 00:07:35.579 --> 00:07:37.980 and then cleaning it up. which, let's be honest, 00:07:38.259 --> 00:07:41.139 often takes the most time. The glamorous part. 00:07:41.620 --> 00:07:43.819 Yeah. Then you do exploratory data analysis, 00:07:44.040 --> 00:07:46.220 EDA. That's where you're really trying to understand 00:07:46.220 --> 00:07:48.480 the data, like listening to its story, finding 00:07:48.480 --> 00:07:51.839 patterns. OK. Data cleaning, EDA. Then feature 00:07:51.839 --> 00:07:54.620 engineering. This is kind of the art form. Turning 00:07:54.620 --> 00:07:57.160 raw data into signals the model can actually 00:07:57.160 --> 00:07:59.339 learn from. Sometimes the biggest improvements 00:07:59.339 --> 00:08:01.759 come from clever feature engineering. So crafting 00:08:01.759 --> 00:08:04.459 the inputs for the model. Right, then you finally 00:08:04.459 --> 00:08:06.579 get to selecting and training in the model, but 00:08:06.579 --> 00:08:09.920 it doesn't stop there. Evaluation. Rigorous evaluation 00:08:09.920 --> 00:08:13.420 and tuning. Using the right metrics, techniques 00:08:13.420 --> 00:08:16.220 like cross -validation to make sure it's not 00:08:16.220 --> 00:08:19.439 just luck. And finally. Interpretation and presentation. 00:08:19.980 --> 00:08:22.240 You have to be able to explain what the model 00:08:22.240 --> 00:08:25.240 does and what the results mean, especially to 00:08:25.240 --> 00:08:27.480 people who aren't technical. That's crucial for 00:08:27.480 --> 00:08:30.019 actually having an impact. That's a great overview. 00:08:30.180 --> 00:08:32.460 It really shows the model training is just one 00:08:32.460 --> 00:08:35.120 step in a much bigger process. So once you've 00:08:35.120 --> 00:08:37.600 got that solid ML foundation, then you're ready 00:08:37.600 --> 00:08:39.379 for the really complex stuff. Then you're ready 00:08:39.379 --> 00:08:42.279 to climb to the summit. Yeah. Deep learning and, 00:08:42.320 --> 00:08:45.019 more recently, large language models or LLMs. 00:08:45.259 --> 00:08:47.320 And deep learning excels where classic ML might 00:08:47.320 --> 00:08:50.059 struggle. Typically, yeah. Especially with really 00:08:50.059 --> 00:08:52.580 complex unstructured data. We're talking images, 00:08:52.779 --> 00:08:55.360 audio, natural language text. The sheer complexity 00:08:55.360 --> 00:08:58.070 there often requires deep neural networks. Before 00:08:58.070 --> 00:09:00.809 we get into specific types of deep learning models, 00:09:01.029 --> 00:09:03.750 are there core concepts everyone needs? Oh, absolutely. 00:09:03.950 --> 00:09:05.629 You have to understand the basics first. What's 00:09:05.629 --> 00:09:08.029 a neuron? How do activation functions work? What 00:09:08.029 --> 00:09:10.450 is back propagation, really? And importantly, 00:09:10.870 --> 00:09:13.309 common problems like vanishing or exploding gradients. 00:09:13.690 --> 00:09:15.769 Why are those gradient problems so important? 00:09:16.370 --> 00:09:19.299 Because if they happen... your model just stops 00:09:19.299 --> 00:09:21.620 learning, or the learning goes haywire. Understanding 00:09:21.620 --> 00:09:24.620 them is key to training deep models effectively. 00:09:24.860 --> 00:09:27.000 OK, got the fundamentals. Then we get into specialized 00:09:27.000 --> 00:09:29.259 models. Right, like convolutional neural networks, 00:09:29.679 --> 00:09:32.519 CNNs. You can think of them as the eyes of AI. 00:09:32.940 --> 00:09:34.940 They're specifically designed for processing 00:09:34.940 --> 00:09:38.379 grid -like data, especially images. OK, CNNs 00:09:38.379 --> 00:09:41.039 for vision. What about sequences, like text or 00:09:41.039 --> 00:09:44.080 time series? That's where recurrent neural networks, 00:09:44.340 --> 00:09:48.730 RNNs, and their More robust cousins like LSTMs 00:09:48.730 --> 00:09:51.029 come in. They have a kind of memory designed 00:09:51.029 --> 00:09:53.110 to handle sequential information where order 00:09:53.110 --> 00:09:56.590 matters. So CNNs for space, RNNs for time, basically. 00:09:56.789 --> 00:10:00.330 Kind of, yeah. But then everything changed. The 00:10:00.330 --> 00:10:02.350 transformer. The transformer architecture. It 00:10:02.350 --> 00:10:04.690 really, truly changed the game. Its core innovation, 00:10:05.169 --> 00:10:07.809 the attention mechanism, was revolutionary. Why 00:10:07.809 --> 00:10:10.330 was attention so different? Well... Older models 00:10:10.330 --> 00:10:13.289 like RNNs struggled with long sentences or sequences. 00:10:13.409 --> 00:10:14.850 They'd kind of forget the beginning by the time 00:10:14.850 --> 00:10:17.330 they got to the end. Attention allows the model 00:10:17.330 --> 00:10:19.669 to focus on the most relevant parts of the input 00:10:19.669 --> 00:10:23.129 data, regardless of distance. And crucially, 00:10:23.570 --> 00:10:26.350 it allowed for parallel processing in a way RNNs 00:10:26.350 --> 00:10:29.070 didn't. And that enabled? That enabled the massive 00:10:29.070 --> 00:10:32.649 scaling that led directly to today's large language 00:10:32.649 --> 00:10:36.830 models, LLMs, all the big ones. GPT, Claude, 00:10:36.970 --> 00:10:38.929 Gemini, they're all based on the transformer. 00:10:39.279 --> 00:10:42.580 It let them understand context across huge amounts 00:10:42.580 --> 00:10:46.039 of text. So when we talk about using LLMs Practically, 00:10:46.080 --> 00:10:48.279 it's more than just chatting with them right 00:10:48.279 --> 00:10:50.519 in a business or research context Oh, yeah way 00:10:50.519 --> 00:10:52.659 more you get into things like prompt engineering 00:10:52.659 --> 00:10:55.039 Which is like the art and science of crafting 00:10:55.039 --> 00:10:57.320 the input text the prompt to get the best possible 00:10:57.320 --> 00:10:59.840 output from the model It's about guiding it effectively. 00:10:59.840 --> 00:11:02.340 Okay prompting is key. What else fine -tuning? 00:11:02.600 --> 00:11:04.960 This is where you take a pre trained general 00:11:04.960 --> 00:11:07.299 LLM and you teach it specialized knowledge for 00:11:07.299 --> 00:11:10.409 a specific task or domain Like medical information 00:11:10.409 --> 00:11:13.169 or legal documents. Exactly. And you often use 00:11:13.169 --> 00:11:15.750 efficient techniques like LoRa, low -rank adaptation, 00:11:15.990 --> 00:11:18.110 so you don't have to retrain the entire massive 00:11:18.110 --> 00:11:20.370 model from scratch. It makes specialization much 00:11:20.370 --> 00:11:22.590 more feasible. Interesting. Anything else on 00:11:22.590 --> 00:11:25.330 the practical LLM front? Yeah, Retrieval Augmented 00:11:25.330 --> 00:11:28.169 Generation, or AIRRAC. This is super important. 00:11:28.330 --> 00:11:31.149 It's about combining the LLM's general knowledge 00:11:31.149 --> 00:11:34.289 with specific external information. How does 00:11:34.289 --> 00:11:37.019 that work? You basically give the LLM access 00:11:37.019 --> 00:11:39.259 to a database, maybe your company's internal 00:11:39.259 --> 00:11:42.220 documents or recent news articles. When asked 00:11:42.220 --> 00:11:44.919 a question, it first retrieves relevant info 00:11:44.919 --> 00:11:47.639 from that database and then uses its language 00:11:47.639 --> 00:11:50.179 skills to generate an answer based on that retrieved 00:11:50.179 --> 00:11:52.899 info. Ah, so it grounds the answer in facts, 00:11:53.159 --> 00:11:55.980 hopefully reducing made -up stuff or hallucination. 00:11:56.299 --> 00:11:58.960 Precisely. And it can provide citations. pointing 00:11:58.960 --> 00:12:01.519 back to the source documents. It makes LLMs much 00:12:01.519 --> 00:12:04.220 more reliable for factual tasks. OK, so you've 00:12:04.220 --> 00:12:07.080 navigated programming, math, ML, even deep learning 00:12:07.080 --> 00:12:09.220 in LLMs. You've built a model in your notebook, 00:12:09.399 --> 00:12:11.679 but that's not the end, is it? Not even close, 00:12:11.840 --> 00:12:14.700 usually. That brings us to that crucial distinction. 00:12:15.620 --> 00:12:19.039 A model in a notebook is a prototype. A model 00:12:19.039 --> 00:12:21.419 in production is a product. Right. And the bridge 00:12:21.419 --> 00:12:23.860 between those two? That's AI engineering and 00:12:23.860 --> 00:12:26.669 MLops machine learning operations. This is all 00:12:26.669 --> 00:12:29.629 about the practicalities of making AI work reliably 00:12:29.629 --> 00:12:33.570 in the real world, solving problems like response 00:12:33.570 --> 00:12:36.269 time, cost, making sure it doesn't crash, monitoring 00:12:36.269 --> 00:12:38.490 it, things you don't worry about much in a notebook. 00:12:38.809 --> 00:12:42.009 So what does the MLOPS lifecycle look like? It 00:12:42.009 --> 00:12:45.149 involves several key stages. First, experiment 00:12:45.149 --> 00:12:47.759 tracking. You need to meticulously log everything 00:12:47.759 --> 00:12:50.980 you try. Parameters, code versions, data used, 00:12:51.220 --> 00:12:53.980 results. Tools like MLflow or weights and biases 00:12:53.980 --> 00:12:55.639 are common here. Keeping track of what works 00:12:55.639 --> 00:12:58.120 and what doesn't make sense. Yeah. Then packaging 00:12:58.120 --> 00:12:59.799 into deployment. How do you actually get your 00:12:59.799 --> 00:13:01.899 model out there? Often involves using Docker 00:13:01.899 --> 00:13:04.320 to create consistent environments and maybe Kubernetes 00:13:04.320 --> 00:13:06.220 to manage running it at scale. Okay, getting 00:13:06.220 --> 00:13:09.620 it running, then what? Automation. Setting up 00:13:09.620 --> 00:13:13.299 pipelines, CICDCT, continuous integration for 00:13:13.299 --> 00:13:16.240 testing code changes, continuous deployment for 00:13:16.240 --> 00:13:19.080 pushing updates smoothly, and continuous training 00:13:19.080 --> 00:13:21.860 for automatically retraining the model when new 00:13:21.860 --> 00:13:24.039 data comes in. Keeping it up to date automatically. 00:13:24.440 --> 00:13:27.220 Smart. And finally, monitoring and observability. 00:13:27.799 --> 00:13:29.940 You need to watch it constantly once it's live. 00:13:30.279 --> 00:13:32.679 Not just system stuff like CPU usage or latency, 00:13:33.059 --> 00:13:36.049 but also model performance. Is its accuracy dropping? 00:13:36.529 --> 00:13:38.629 Is the input data changing unexpectedly? What 00:13:38.629 --> 00:13:41.149 called data drift? Is it becoming biased? So 00:13:41.149 --> 00:13:43.429 it's an ongoing process, not just a one -time 00:13:43.429 --> 00:13:46.009 deployment. Absolutely. Production AI needs continuous 00:13:46.009 --> 00:13:48.190 care and feeding. Wow. Okay. That's a really 00:13:48.190 --> 00:13:51.149 comprehensive roadmap. Programming, math, ML, 00:13:51.309 --> 00:13:54.090 deep learning, ML Ops. Yeah. It leads us to the 00:13:54.090 --> 00:13:56.669 final piece, right? The capstone. Exactly. Building 00:13:56.669 --> 00:13:59.210 your project portfolio. Because at the end of 00:13:59.210 --> 00:14:01.309 the day, all this knowledge is kind of theoretical 00:14:01.309 --> 00:14:04.210 until you apply it. Sure. Don't just tell. Precisely. 00:14:04.429 --> 00:14:07.029 A portfolio of solid, well -documented projects 00:14:07.029 --> 00:14:09.230 is the single most powerful way to demonstrate 00:14:09.230 --> 00:14:11.840 your skills. much more than just listing courses 00:14:11.840 --> 00:14:14.399 you took. It shows what you can actually do. 00:14:14.820 --> 00:14:18.059 So what kind of projects should people aim for? 00:14:18.179 --> 00:14:20.320 Does it vary by skill level? Yeah, definitely. 00:14:20.399 --> 00:14:22.679 If you're just starting out, maybe analyze a 00:14:22.679 --> 00:14:24.980 classic data set, something from Kaggle, focus 00:14:24.980 --> 00:14:27.519 on the storytelling, good visualizations, clear 00:14:27.519 --> 00:14:30.200 explanations, maybe a simple baseline model. 00:14:30.299 --> 00:14:32.539 Okay, for beginners, what about intermediate 00:14:32.539 --> 00:14:35.000 folks? You could try replicating a published 00:14:35.000 --> 00:14:38.409 research paper. A simpler one, maybe. Or build 00:14:38.409 --> 00:14:40.830 a complete end -to -end web app that uses an 00:14:40.830 --> 00:14:43.289 AI model you trained. Show you can connect the 00:14:43.289 --> 00:14:46.590 pieces. And for more advanced learners. Go deeper. 00:14:47.190 --> 00:14:49.730 Fine -tune an LLM for a really specific niche 00:14:49.730 --> 00:14:53.409 task, like that legal chatbot idea. Or build 00:14:53.409 --> 00:14:56.190 out a full MLOPS pipeline for one of your projects, 00:14:56.610 --> 00:14:59.090 showing you understand the production side, demonstrating 00:14:59.090 --> 00:15:01.370 that whole lifecycle. Those are great examples. 00:15:01.529 --> 00:15:04.480 And once you build something... What's the best 00:15:04.480 --> 00:15:07.139 way to showcase it? Don't hide it. Put it on 00:15:07.139 --> 00:15:09.419 GitHub. That's standard practice. Make sure the 00:15:09.419 --> 00:15:12.519 code is clean and write really clear documentation. 00:15:12.620 --> 00:15:15.240 It's a good readme file explaining what it is, 00:15:15.259 --> 00:15:17.159 how to run it, what you learned. Documentation 00:15:17.159 --> 00:15:20.919 is key. Crucial. And a pro tip, write a blog 00:15:20.919 --> 00:15:23.700 post about your project. Explain your process, 00:15:23.840 --> 00:15:26.700 the challenges you hit, how you solve them. It 00:15:26.700 --> 00:15:29.240 solidifies your own understanding and it acts 00:15:29.240 --> 00:15:31.559 like a living resume for potential employers 00:15:31.559 --> 00:15:34.179 or collaborators. That's really practical advice. 00:15:34.419 --> 00:15:36.039 Well, we've certainly covered a massive amount 00:15:36.039 --> 00:15:38.659 of ground in this deep dive on becoming an AI 00:15:38.659 --> 00:15:41.659 builder. It's a lot, for sure. And the key thing 00:15:41.659 --> 00:15:45.240 to remember is that this journey, it's a marathon, 00:15:45.519 --> 00:15:47.620 really. There's no finish line because the tech 00:15:47.620 --> 00:15:50.759 itself is always changing, always evolving. That's 00:15:50.759 --> 00:15:53.039 a bit daunting, isn't it? It can be, but the 00:15:53.039 --> 00:15:55.179 good news is the foundational principles we talked 00:15:55.179 --> 00:15:57.460 about, solid programming, understanding the math, 00:15:57.860 --> 00:16:00.600 thinking in terms of systems. Those things don't 00:16:00.600 --> 00:16:03.139 change much. Ah, so those are the enduring skills. 00:16:03.299 --> 00:16:04.879 Exactly. They'll serve you well no matter how 00:16:04.879 --> 00:16:07.299 the AI landscape shifts. They're the constants. 00:16:07.480 --> 00:16:09.899 So the roadmap isn't just a checklist to tick 00:16:09.899 --> 00:16:12.080 off. No, think of it more like a continuous cycle. 00:16:12.240 --> 00:16:14.960 Learn something new, then build something with 00:16:14.960 --> 00:16:17.679 it, then iterate based on what you learned. Repeat. 00:16:18.179 --> 00:16:20.919 Learn, build, iterate? I like that. Start with 00:16:20.919 --> 00:16:22.600 the foundations, keep building projects, and 00:16:22.600 --> 00:16:25.820 just never stop learning. That's the core philosophy, 00:16:26.220 --> 00:16:29.220 yeah. It's challenging, no doubt, but also incredibly 00:16:29.220 --> 00:16:31.740 rewarding when you see it come together. So maybe 00:16:31.740 --> 00:16:33.639 a final thought for everyone listening. Given 00:16:33.639 --> 00:16:37.639 how fast AI is evolving, how can you adapt your 00:16:37.639 --> 00:16:39.259 own learning? How do you make sure you're not 00:16:39.259 --> 00:16:41.200 just keeping up but actually building with it? 00:16:41.340 --> 00:16:44.419 Yeah, and maybe reflect on today's roadmap. What's 00:16:44.419 --> 00:16:46.720 the one foundational step you could prioritize 00:16:46.720 --> 00:16:49.460 next? What's your next move to get out of that 00:16:49.460 --> 00:16:52.360 potential tutorial hell and really start building?