From .NET to AI 🚀 | What is AI, History, Types, Branches & Machine Learning | Session 1

Vídeo em Inglês, publicado pelo canal RI-TECH by Achyut Kendre.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video provides an overview of artificial intelligence, machine learning, and deep learning concepts, history, and applications, while outlining a comprehensive training curriculum focused on implementing these technologies using C# and .NET instead of Python.

Key points

• Artificial intelligence is defined as a branch of computer science that creates systems capable of performing tasks requiring human-like intelligence, encompassing perception, learning, reasoning, problem-solving, and decision-making.
• The history of AI evolved from Alan Turing's 1950 theories and the 1956 Dartmouth conference to modern developments in large language models, multimodal systems, and agentic AI.
• Machine learning differs from traditional programming by utilizing historical data to automatically identify patterns and make predictions, rather than relying on explicitly coded rule-based logic.
• Machine learning is categorized into supervised learning, unsupervised learning, reinforcement learning, semi-supervised learning, and self-supervised learning.
• The presented training program covers AI fundamentals, ML.NET, deep learning, natural language processing, transformer architectures, vector databases, RAG, agentic AI, and enterprise integration using C# and .NET.

Topics

Artificial IntelligenceMachine LearningDeep LearningNatural Language ProcessingNeural NetworksSupervised LearningReinforcement LearningRAG

Video transcript

Transcript from available captions; it may contain errors.

[music] [music] is a branch of computer science that aims to create machines or softwares that can perform the task that normally require human intelligence. In other words, artificial intelligence in the field of Kimrasan that focusing on creating systems capable of performing task that normally require humanlike intelligence. So AI is very wide branch. AI is not a one or two things. It's a computer science field where we design develop test deploy implement systems. Those are capable to perform the tasks normally require humanlike intelligence. So think of this way when human sees a picture of a cat so we instantly know it's a cat. When we speak, we form sentence. when we drive so we make split second decisions. AI tries to make machines do these same things. The goal or purpose of the AI to make the systems of intelligent like human. What are the core capabilities the human capability they try to replicate? So the first good capability so AI tries to replicated learning improving from experience of data. Second, third will try to reason the reasoning for drawing the conclusions logically. Third, problem solving, finding solution to challenges. Four perception. Understanding images, sound, text. Fifth, language understanding. Understand and learn or generate speech and text decision making choosing the best action. So you should remember so AI is not just a robots it's a broad field a chartboard a Netflix recommendation engine self drawing cars spam filters are all AI so mathematically if you want to remember so what is AI so AI is making machines think act intelligently like humans Fine. Now this include so perception, understanding, learning, reasoning, decision making and taking the actions. Now we got an AI. So AI is the broader field of computer science with whose goal or whose objective is to create the systems or design or develop such a systems to replicate the human intelligence. They can perform the task just like humans. Now let's go to the history of AI. So AI didn't appear overnight. It evolved through decades with exciting heights and disappointing AI winters we can say. So if you look at to the major milestones in AI, so it just started in 1950 when Alan Turing put the theory that can we create can we have a branch where we can study to create intelligent systems or intelligent machines. in 1950. In 1956, so we got Dartmouth conference where many scientists, researchers, university professors gathered together to decide can we have one separate branch nowadays we call it a IML to study, research and further. So how we can pro proceed further? and how we can take it to the next level. And the result of this is in 1980 we got the experts and AI got the limelight in 1997 when IBM DW beats the Gary Kpares. And in 2012 we got the deep learning neural networks. The human like chain of neutrons or perceptrons which learn from the pattern called alexet. So initially it's called Alexet and it change the face of because this deep learning when combined with attention mechanism and transformer architecture. So you got the lens large language models and in 2020 so the first commercial model we got is GTP3 and then you find many models in 2023. So we got the GTP4 and charge GPT and with the release of charge GPT so AI is not a new thing. So AI is everywhere. So in 2025 now we have the multimodal agentic AI generative AI. So LM is the example of generative AI. So agent PKI multimodal system expert system we are getting many many many more things done. So it's not happen overnight. So it's the effort of decades from 1950 to 202526. So AI progresses in cycles. Hype, disappointment breakthrough hype again. We are currently in massive food. Now why many we can say AI researchers or big tech company owners or techs are worried about AI due to speed. what they were not able to achieve from 1950 to 2012 that achieved it from 2023 to 2025. So within we can say last three or two here two years there dramatic changes in AI and AI is learning very fast. So it's learning very fast, right? It's learning very fast. In previous sessions, we had a word about this. So in last session, so you don't need to worry about it. So it's not going to replace you. AI is not going to replace you. The person who is able to use the AI who know the detail know who has a detailed knowledge of AI will replace it. Now AI programming is little bit different from the traditional programming that we were doing from many years in traditional programming. So what we do we have the rule based code. So we write the rules and your code uh sorry in the code and when you run that code it follows those rules. So it will perform as per those rules. So the ML and AI is different. It's not like your traditional programming. For example, suppose we want to determine whether the transaction is forant. So in traditional programming, so what we do so that we do in a net or java or any other language. So we put some set of rules together like if amount is greater than one lakh and the country is not equal to customer country and transaction time equal to unusual then fraud. So we write the set of rules we write these rules. So today so we are doing this this kind of programming traditional programming. So conceptually we write the rules in the data in a program and then it will output it. Write the rules and the data in the program and it outputs it. Now this programming got changed in ML and AI. In ML and AI we don't write any rule. So we don't write any rule based or logic based coding or logic based programming. So there is no need of logic. ML and AI it takes a different approach. So instead of writing logic and the condition if this then do this as this then do this repeat this for 10 times or do this for five times. No we are not going to do that in ML or EA. except that we provide the historical data huge data to the ML and AI your ML and AI program and from the datas ML and AI automatically find out the patterns and based on that pattern it will do the task for you. So based on that pattern it learn the things and that learning will help you to execute some task or activity. So for example suppose we have a list of transactions like transaction one transaction two transaction three transaction four transaction five like this. So suppose we have historical data of 10,000 transactions where uh we had already identified this was the fraud transaction this was genuine transaction this fraud transaction this was genuine transaction. So if you provide the data to the EI and ML program what it does it learns the pattern from that particular data. So ML algorithm we call it is in model. So learns the pattern. So it learned the pattern. Okay. So we have the historical data. Then that historical data will be get converted to ML algorithm. So we'll provide that historical data to the ML. Sorry. So we'll provide the data. So we have 10,000 records of the transaction. So we provide that data to ML algorithm. So ML algorithm will learn the pattern from it. So we call it in so when you learn the patterns we call it in train model. So it will get you the trained model and when you provide a new transaction detail to that particular model it will predict based on the learning whether that particular transaction the fraud orient. So this is the major difference fundamental difference between the traditional coding or MLA programming or MLA coding. In MLAI we do not code logic, we do not code things like this. So in ML so we train the model on historical data and from that it will learn the patterns and the learned patterns can be used to do the future predictions or to do any kind of other task also. So in this example we are doing just a prediction whether the transaction is product or transaction is genuine. So this is the major difference. Now let's have a look at two types of EI. AI can be classified into two ways. So we can divide the AI into two uh we can divide the AI by two ways. So the first way is based on capability. So based on the capability, so if you divide the AI based on the capability, so it will have the we can say three types. So the first is narrow AI. So we can say so narrow AI or weaki performs one specific task. Well, so narrow AI or weak AI designed to perform only one task specific task. So we create a AI program or application whose responsibility just to classify the tickets. So support ticket. So it will just classify the ticket. It's not able to do anything other than that. So such kind of AI is called narrowi. It's a weaki. It cannot generalize beyond that task. For example, Siri, Google translate, Netflix recommendations, spam filter. These are all examples of narrowi. We are using narrowi from many years. So we are already using this. These are already exist. Now general AI so that we are using from 20 uh you can say 20 24 or you can say 23. So general AI that support strong AI or EGI it's a hypothetical machine with the human level intelligence across all the tasks. So general AI is a hypothetrical AI with human level intelligence and it can perform any any task. So any unlike task can learn, reason, apply knowledge to any problem like humans. So technically people says it does not exist eight. This is what researcher chasing. But what you are looking around yourself like JPT, Gemini, cloud or many other we can say AI models. So when you look through those models, you'll feel we have general AI right now. So narrow is now history and we are using general AI because it's chart GP or GP model or cloud or Jim they are able to do many task. They can answer the question they can write the essays they can reply the mails they can create generate the images they can format your content they can uh solve the quiz they can uh help you in many many areas. Still there are few areas that is unexplored but it's going to happen within one or two years. So right now we can say we have the general AI or strong AI or AGI and next thing that computer scientists or researchers are trying to achieve is super AI the super intelligent AI and uh if you look at last 3 years so you might realize the AI is learning very fast. So AI is learning very fast. Super AI is hypothetical AI that surpasses human intelligence in every domain. would be able to improve itself purely theoretical often discussed in science fiction and ethics debates but the way the AI learning the things I don't think so okay it is unactable so you'll find the super intel AI within years or within we can say 5 to 10 years Now these are the three categories of the AI based on the capability. Now if you divide the AI based on the functionality so we can have the four types of AI. So reactive machines no memory response only to current input. Example IBM deeply so it can play only chess. They respond based on the various moves or happened in a chess game. It's limited memory uses past data for short time to make decisions. So examples are self-driving cars. So theory of mind to understand the emotions, beliefs, intentions of others doesn't fully existed being researched self-power. So has consciousness and self-awareness purely hypothetical. So there are the four types of AI based on the functionality reactive mach limited memory AI the area of mind AI self-paw now look at let's look at in which which are the major branches of AI so AI is a huge umbrella here we are the main branches the first and the most important branch branch mention that the fundamental branch on the top of that all other branches are cultivated we can say machine learning ML and in ML so system learns from data instead of being explicitly programmed just before I told you so here we do not do any explicit programming so we do not write any logic or looping or so we just provide the data so your program will learn from the data and it will do the things for like predicting house prices or predicting share prices or predicting whether the transaction genuine or not. Predicting whether the image is a cat image or a dog image or predicting anything or computing the price of house. The second branch is deep learning. So first we got the ML after ML we got the neural network and neural network gives the birth to a deep learning. So DL and DL is also called so modern ML and DL uses the concept of neural network just like a human brain. So human brain contains trillions of neurons and every neuron is connected with another neuron and that neuron has the intelligence and the memory. So what computer scientists did they tried to replicate the human brain working with the help of artificial neural networks. So ANN that we are going to see and that gives birth to a next step in ML called deep learning. Deep learning we can say model ML a sub field of ML using multi-layered neural networks and it can do face recognization. So ch it can do many things. It can scan the images. It can identify the images. It can identify the movements more intelligent than ML. Then this deep learning combined with self attention mechanism which is the backbone of a transformer architecture use NLP. So this ANN or LSTM or GANs were having some problems and those problems are sorted using one mechanism called the self attention. We learn all these things in detail. So Google had introduced one mechanism called self attention and self attention gives birth to a transform architecture and the transform argue gives birth to NLP. So next branch of AI is called natural language processing NLP. So teaching the machines to understand and generate human languages. This NLP in NLP so we try to create systems or application those are capable to understand human language. If you look at to the charge now charge GPT uses NLP internally because it can understand many languages like Hindi, Marathi English French Russian. So NLP was there with us from many years with a product called Google Translate, Cardboards, sentiment analysis. Then we have the next branch of AI called computer vision. So making machines see and interpret the images and videos. Now this has very vast you can say applications in computer vision branch. So your AI or systems or machines can see the images in video and they can interpret what is happening there. Example, self-driving cars, medical imaging, facial recognization. Now the next thing that we have is robotics. So designing the intelligent robot that perceive and act in physical world. Now up to computer vision. So your AI enamel is inside your machine without any kind of hands, legs or any kind of movement but with the help of robotics. So your AI and ML will be available in your physical world. They can act do the stuff like humans. Example we are already using robotic arms and factories humanite robots. Now the next thing that we have is export systems. So program that mimic a human expert decision making using rules like medical diagnosis detection speech recognization and synthesis converting the speech to text like Alexa Siri voice to text planning and scheduleuling so making we can say sequences of action to achieve the goal logistics root planning Filo. So dealing with partial tooth not just ur one washing machine air conditioners they use a fill logic. So genetic sorry genetic algorithm evolution computing problem solving inspired by sorry inspired by the biological evolution optimizing network game strategies. So in directly you can say so D is a subset of ML and ML is a subset of EI. EI is the biggest circle. ML is inside it and D is inside. Now let's explore machine learning. So let's have some fun about the machine learn. Let's define the machine learning by Tom Michael. Machine learning is a computer program is said to learn from experiences with respect to some task and perform measures. If it performs at task in T are measured by P improve with experience belief. Let's redefine it's the computer program that is said to learn from experience B with respect to some task D and perform major field. a program that learns from previous data or previous experiences for specific task that you want to do and perform that particular task for you and it performance at a task in T is measured by P improve with experience E. So as it experience goes so its performance is going to improve to perform that particular task to your task. In simple words, ML is subset of AI where machines learn patterns from data and improve with experience without being explicitly programmed for every rule. So we are not going to program them explicitly. Fine. So ML is basically a subset of AI. So where machines learn patterns from data and improve the experience without being explicitly programmed for every rule. Now traditional programming versus machine learning that I already told you. So in traditional programming so human writes the rule. So in machine learning so machine find the rules from data. So we don't tell them we don't uh tell any kind of rules. Now in traditional programming we input data plus rules together and it will generate the output. In machine learning we just input the data and what output we are expecting it will figure out the rules on its own. It will learn the rules on its own. Now hard to handle complex task wait for complex patterns. So how to handle complex task great for complex pattern. Now analogy. So what we have here? So traditional programming is like giving a child a rule book. So if red light stop, if green go. ML is like showing the child 1,000 traffic videos and letting them figure out the rules themselves. So for example to retake the spam we don't write rules for every spam word. So you feed the model thousands of spam and non-spam email. It learn the pattern itself and it will uh classify your email to a spam or not spam pattern. Now why machine learning? So why not just write the rules manually? Hence why ML is essential. Some problems are too complex for manuals. Suppose if I tell you to write a program in C or asnetate to recognize the face, you may say it's impossible because there are many many many rules, thousands, lacks of rules. In fake regression, you cannot write a rule for every phase. Similar example, language translation. The grammar rules are endless and messy. So, ML is necessary for some problems. Those are too complex for manual books. Now, when data is growing exponentially, so we need ML. So, we generate billions of data point daily. So, ML on the data. So when you have huge data your rule based programming is going to fail or it will be very slow. So we use the ML adaptability. So ML models improve as more data arrive. No static rules. So your rule based coding or classical or traditional coding cannot improve on its own. So if you want to improve you need to change the code. You need to put new rules. you need to modify the existing rules. It's not the case with the ML. So ML model improves as the new data arrive you feel more and more new data to the ML. So it will improve on its own hidden patterns. So ML find the patterns humans can't see hidden patterns. I will find the patterns human can't see. For example, customer buying behavior, automation at scale. So we train an ML model can make millions of decisions instantly. So scaling is again one reason. So we go for ML. The fourth is also very important. So sometimes as a human we can't see many patterns or many things but ML can find out. ML may surprise you by finding such patterns or such learning that even humans could not think. Sixth is personalization. Netflix, Amazon, Spotify all use ML to personalize for each user. So we cannot write rulebased coding or traditional coding per for every person. The personalization is very challenging in your rule based or traditional coding. Not only the six, there are many other reasons but there are the six main regions or sorry reasons that's why we need ML to the bottom lines. So ML allows us to solve the problem that are impossible or impractical [clears throat] to solve with handcoded rules. Now let's have the types of machine learning. So we are going to study the machine learning in detail. No problem. So there are three main types. Let's do hybrid ones. So one is supervised learning. We'll see what is supervised learning coming in sessions. So in short I'll tell you the supervised learning. So your machine ML learns under the guidance of human. So human will provide the input and the output both correct output here before uh in previous example. So if you want to identify whether a transaction is important or genuine what we do we provide the data input of all the transaction labelled with this is transaction this is genuine this is problem this is genuine this is fraudent. So when we provide such kind of data we are training the machines under our guidance. So we are telling to the machines this is the pent transaction and this is a genuine transaction. So such kind of learning is called supervised learning. What is the goal of supervised learning? So learn the mapping from input and output. It will map the input and output and it will find out the pattern based on the input and output. So analogy a teacher gives you questions with answers you learn from that. So in simple layman language you can say so teacher will provide you the questions and answer books and you learn from that question and answer supervised learning. So two subtypes of supervised learning that we have the classification regression we'll see that. So classification will help you to predict a category, spam, not spam, cat, dog. Regression, it will help you to predict the number like house price, temperature, share price, fine, sales forecast for next year. So algorithms, so many algorithms are there in supervised learning. So linear regression, logistic regression, decision tree, SPM, random forest, neural networks. So we see a few of them now examples. So email spam detection, image classification, price prediction. These are the examples of supervised learning. Unsupervised learning. So we provide this data without answers. So for example, we provide the list of transactions without a label that tells that this transaction is important or this transaction is genuine. So we provide the data to train your we can say ML such kind of training it's called unsupervised learning in unsupervised learning so your ML will find hidden structures and patterns right now for example in lemon language we can say suppose if you are given a box with a mixed fruit and ask you to group them no one tells you the categories is so you can group the fruits based on your learning. So maybe few of you will group it based on the test. You may me group them based on the color. You may group them based on the vitamins and we can say content of that particular fruit or location or season. We cannot predict how it grow but it will grow uh that predicting based on the structures and the patterns. So unsupervised learning will not tell you as no or some value. It tells you the category and that category is not provided previously. So it will figure out the category based on its own. There are two type subtypes of the you can say unsupervised learning clustering and association. So clustering will help you to group similar items and association will find the rules. People who buy bread and also buy the buttons or people who buy the laptop may buy the mouse or USB pen drive. The algorithm that we have here came in hierarchical clustering divori PCA. So we have many algorithms. So we'll have few of them. Examples of the we can say unsupervised learning are market based analysis, analy detection, topic modeling. Now the third thing that we have is third learning is reinforcement learning. RL in RL or reinforcement learning. So we do not provide any data. We do not provide any input. We do not provide data. We just provide the environment and we'll tell the ML this is the environment. Do whatever you want to do. Interact with this environment and learn from it. So the goal is maximize reward over time through trial and error. So for example, so training a dog with a taste like good behavior, reward, bad reward. Now key concepts are agent, environment, state action reward policy. The algorithms used are Q-learning, deep network, DQN we can say in short PO. So the third type of ML examples are playing AI, AlphaGo, robotics, self-driving cars. Now there is one more category. It's not a direct category. Direct categories we have only three. So supervised, unsupervised and reinforcement. So few people say semi-supervised learning is again one of the learning technique ML learning technique. Here we have mix of small label plus large unlabelled data. So I'll do the mix. So we'll provide few transactions with labels like fraud, genuine and few transaction without any label. Like Google photos you label few faces. it learned the rest. Now self-supervised learning and this is the uh we can say thing due to many AI researchers are worried self-s supervised learning. So model creates it own labels from data structures. So model model means trained algorithm. So algorithm who has learning. Now for example so GPT models. So predict the next word from massive text. We'll see what GP is basically a next word predictor. So if you compare them quickly supervised we provide the label data. Goal predict examples spam detection unsupervised unlabelled data. So goal is not to predict discard the patterns. like customer segmentation reinforcement reward signals. So here no data no label nothing. So we'll just provide the environment and it will do the try and error and based on the success and error we can say failure it will be rewarded like maximize reward the for example game semi-supervised mostly unlict web content classification for example self-supervised atrogenerated so it will generate the data on its own learn the presentations on its code and chart detectively. Now these are the types of ML. Now there are many applications of machine learning. For example, uh in healthcare, disease diagnosis, cancer detection from exit, drug dis drug discovery. So personalized treatment in finance we can have the credit scoring, fraud detection, stock market prediction, algorithmic trading. So in e-commerce and retail so we have the product recommendation Amazon demand forecasting dynamic pricing social media and entertainment so ranking Facebook or instant Instagram so content recommendation Netflix and YouTube Spotify take new detection in transportation we have the self-driving cars traffic prediction route optimization Google maps so we we are already using this. So natural language. So in natural language we have the chatboards, RP translation, Google translate, voice assistant, Alexa Siri, sentiment analysis, computer vision, so face recognization, medical imaging, object detection, autonomous drones, cyber security, so intrusion. So this is one area. So where AI researchers are worried. So they are saying in few years uh countries will develop AI enabled weapons. So the weapons they can fire or operate on their own no human interference required. So cyber security so intrusion detection, malware classification, anomaly detection, agriculture, so crop field prediction, pest detection, preciation farming, manufacturing, so predictive maintenance, quality control, supply chain optimization, education, so personalized learning, automated grading, student performance prediction, the gaming. So here we have the AI opponents, Alphago, chess engines. So procial content generation and many more. So you'll find the AI, ML everywhere. Now if you compare AI, ML, so AI is the broader field making machines intelligent. ML is subset of AI learning from data. Scope in scope of AI is very vast, wide. The scope of ML is narrow inside AI. So goal similar to human intelligence learn patterns to make the predictions. Now in AI we have the rules logic search learning it is everything combined together. So here we have statistical learning for data. So ML based on learning from data. Now data dependency may or may not need data. So always need data. So ML needs data to learn. Now examples export system robotics NLPML. So spam filter recommendation engine. Now relationships of parent and this is child fit with them. Now if you go to the key analogies so is like transportation. So huge category cars, trains, planes, bicycles, ML is like car. So one specific powerful type of transportation. So all cars are transportation but not all transport is car. Similarly or ML is AI but not all AI is ML. Now another thing that you need to remember. So rulebased chess engine that plays by hardcoded rules. So AI not ML. a chess engine that learns by playing thousands of games that is AI plus ML. Now deep learning. So deep learning is subset of ML and ML is subset of EI. So we are going to learn all of this. So inside characteristic now this for today. So before closing so I will show you. So what we are going to cover throughout this brick training. So topic wise. So the first thing that we are going to do is we are going to build a and ML foundation. The first thing that we are going to do is we are going to build a IML foundation. So in AL ML foundation. So first thing that we'll do is AI fundamentals and ML. We'll learn the AI fundamentals in ML here. So few of them we had learned right now. So what is AI is the traditional programming types of branches MLY ML types of ML now then tomorrow so many of you don't know the net 10 so tomorrow we'll cover the foundation for the net 10 and there we cover the modern C# a 4 project uh structure EF core middleware dependency injection reporter pattern the code operation just to give a hands on uh darknet core 10 so that uh we could we are going to use net port 10 a port 10 for all the demos you and we are going to the visual studio 26 so we'll have a small session which will help you to upgrade from your uh upgrade to dark net now next thing we'll deep dive into the ML so deep dive into the ML in ML so we'll learn what is models so how to create model we'll have the supervised learning and supervised learning reinforcement modeling in detail their algorithms like regression classification clusting algorithms. So why and what MLN net? So model builder. So we'll use the feature ML.NET framework. So we'll learn the ML.NET framework. So net has a built-in framework called ML.NET which will help you to implement the ML algorithms. Then we'll create the model trend split data. We'll do the prediction. So we'll consume model from the esport. So we'll build a4 application to consume the ML model. So we'll implement many practical demos. So few of them are listed here. So evaluation uh so we'll do the practical demo there. So random forest algorithm we are going to implement logistic regression pass three so like GBM so we implement various examples like we can say spam or not spam or failed or pass or we can say price prediction of a house or price prediction of a share from historical data. So using various algorithms like random force regression, logistic regression, fast tree and light GPI. Now after that we'll go for evaluation matrix. So we'll learn effectiveness of your model. So once you create a model and uh when you train that model so we also need to worry about the performance of that model. So we should learn the effectiveness of that little model. We should learn the effectiveness of that metal model using evolution matrix overfitting underfitting confusion matrix and accuracy appreciation call limitations of classical ML so that we'll cover in ML in detail now after that so we'll go for so deep learning so in a deep learning we'll learn the basics of a deep learning so why what and why deep learning neural networks we going to study neural networks in detail what is perceptron what are the layers or what is backrop propagation, what is the activation function, forward propagation, loss function, so how the neural network learns the things, what are the gates, what are the we can say loss gates and how it optimize itself. So what is vanishing gradient? So many things we are going to learn about the deep learning. Now once we cover the deep learning so we'll go for the uh implementations of the deep learning like CNN. So where we learn the convolution filters polling image concepts RNN and sequence learning vanishing gradient LSTM sort and ONX load existing model. So if you have any existing model pre-built in a Python or Java C++ or any other technology so you can also load that model into your net application using ONX and you can study that model at runtime. You can find the inputs and output from that model and then you fine-tune retrain that particular model. then uh you can find out the internal details of that particular model. Now then we'll have the practical demo like image classification ticket classification project using uh we can say CNN and we use RNN here. So that we are going to cover in uh we can say uh deep learning or neural network of them. Now after that we'll go to the details of a transformer architecture in the top of transform architecture. So first we'll go for NLP fundamentals. So natural language processing parameters there we'll learn the what is NLP text processing vocabulary tokenization top stop wordsing leization engram uh TF coign similarity so including distance man so we have many things here so we'll learn that in NLP so once we learn the NLP then we'll go for we can say modern NLP so where we learn the attention mechanism that the Google mechanism and then we'll learn the transform argue So we'll see the what are the classic limitations of the classical LP attention mechanism self attention cross attention uh we can say entropy methodology so queries keys and values encoder decoder transform architecture that we can say backbone of your modern lens now vector embeddings tensors so uh contractual me embeddings so and GTP then uh we also learned the minialm knowledge distillation knowledge transformer. So teacher child model relationship of that. So that we are going to cover here in NLP. Next thing that we go for embeddings. So we use the mini that the Microsoft we can say model encoder a dense embedding vector representation cosine similarity semantic search that is very important mechanism nowadays every enterprise application needs. So then we'll go for ranking then we'll go for implementation of this and here we'll have the practical demo of implementing the contextual embeddings. So how to generate the contextual embeddings using a mini ML. Now then we'll say we'll go for so LM fundamentals. So in LM fundamentals we'll cover so what can y tokens context window inference generation temperature capabilities and limitations. So what is GTP LM? What is lama or what are the GTP models or feed models. So we'll study about those LM models of that. Then we'll go for the openi and chart gemini integration. So next thing that we'll do we'll integrate the open chart uh we can say chart GPT or Gemini and uh we can say uh will inspe what are the inputs and output is expecting what is the structure of the every component of the gemini similar way we study the openi GPT we can say API with that then we'll integrate it in as code then we'll go for structure output so how to generate structure output how to do the streaming. So sorry streaming then provider abstraction. So how we can use the factory design pattern option design pattern to abstract the provider. So later on if you want to integrate the cloud in same application or any other uh we can say model. So you can do it with very less changes inside the code. Now here we use the options pattern pattern design pattern. Then here we'll do model integration and we'll do various examples. So many examples of that. Now next thing that we'll do is so we'll go for Olama. So we'll use the Olama. So we'll install the Olama. We'll install the local models. We'll download the models and we'll do the same things with the local model also. We'll use that it from the east. Core application or C# application and we'll do the uh many things. So how to store the conversation history or how do we structure output from we can say locally deployed models using Olama and here we'll implement a chart application using Olama and in uh we can say open AI charging we also will implement the two three applications here at the every step we'll implement some application now then go for the advanced LM concept advanced LM so we'll first learn the advanced LM architecture so there we'll learn the LM architecture EI So we say chart API what is the memory how to manage the memory in the context SQL persistence here so how to persist the memory in a SQL server database security observability so how it observe and react to the results generated then we have the conversation context conversation summary practical demo here for advanc and then we'll go for the rules in function calling so where we'll cover MCP in detail so model context protocol so there was see What what is the functional tool? How to call it? What is the schema? What is the external APIs? Database tool, how to call the database tools, control business actions, or session. What is MCP? What is MCP client? How to create MCP client? MCP server, MCP RA. Here we'll do two uh two three demos to understand the tools, functions, calling and MCP that we are going to cover in advanc. Now the next thing that we'll cover is uh agent TKI. So in agentic AI so we'll learn what is agent. So what is uh agentic AI? So agent fundamental concept like memory agent tools agent hands and legs. So instructions promps tools memory everything of them. Here we'll create a agent from scratch using C# and core without using any framework. We are not going to use any framework. So we'll build our agent from scratch in C# and AL course and then we'll use the agentic framework. So then we'll learn the frameworks. There are many agent frameworks like semantic kernel or we can use the lang chain. So Microsoft has latest framework called MF. So we are going to use the MF and this will be the feature framework for creating the agents and the Microsoft platform. So here we'll cover the Microsoft agent framework. So here we'll cover what is MAF. So various framework, semantic kernel, kernel services. Then we'll cover many concept like prompts, plug-in, functions and many things. And here we'll have full page agent and here we'll also implement multi- aent system where multiple agent coordinating agent orator agent and all the agent will do some specific task for that. So here we'll have one fulllex demo implementing the agent using MF. So mixoft agent framework with them then we'll integrate the MCP and rack. So both so MCP and rack to uh we can say so MCP and rack to your agent. So here we'll learn MCP concepts again. So tools, resources, prompts, MCP workflow, agent tool architecture, patrol enterprise integration. So we'll have a practical demo how your agent can use the MCP and [clears throat] how your agent can use the Ara. Now then we'll go for the enterprise AI engineering sorry topic sorry. So vector and the vector in the rag here. So in the vector database and rag we learn the quadrant collections vectors metadata chunking top retrieval hybrid search that is again very uh important feature nowadays every enterprise application needs then we'll see the fundamentals of the rag and we'll implement definment question answer project here using a rag so we'll study the rag in detail so we'll uh do many many concepts in the rag and we'll implement a practical demo based on the rag and the vector db or quantum db be here like doing the code operations incardant and sash plotting and retrieving the embedding semantic search hybrid search document search here we'll implement project where we'll upload the PDFs and we'll ask the question to the AI and AI will generate the answer based on that particular PDF that so based on that particular PDF it will read the PDF understand those PDFs it will store the memory and when we ask the questions so it will answer those uh questions now so in agent after agent AI will go to the enterprise EI. So where we'll have the architecture security evolution grounding health scalability looping guardrails monitoring deployment many many things are there. So how to control the cost, how to optimize the performance, how to deploy the AI applications or agents in Azure AI services. So uh how to build the agents and run it using Azure AI services that we are going to cover here. So we are going going to cover all these things in enterprise AI engine. And at the end so throughout this particular we can say modules we'll implement many uh small small projects and uh throughout we can say so after this we'll at the end we'll implement one large project where we'll do everything from end to end. So throughout uh this project we use the clean architecture. So we'll use various design patterns and uh enhancable architecture and we'll try to utilize all the concepts that we had learned including your MCP rag or we can say agent or identici lms so embeddings so contextual embeddings quadrant hybrid search so we'll combine all of them then we'll deploy it in the azouri and we'll also uh uh can say generated documentation and we'll put it into github so you can add that particular project to your portfolio and then we'll go for the interview questions. So how to prepare for the interview? So that how to prepare for interview. Now interesting thing about this particular uh training that I already told you that you don't need a python so you don't need a mathematics. So everything will be in a C# and a spot with code. So everything will be on a net platform. So entire thing will be on the net platform. Fine. So this is the thing that I have to tell you today. So if you have any question answer sorry questions please ask me.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, AGI, ANI, ASI. Publicado no YouTube em 2026-10-09.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.