hi everyone. In this video, I'm going to show you guys five AI projects that you can get started working on immediately. For each project, there will be three levels. The first one is just going to be prompt engineering and some no-code tools. It's actually pretty crazy how much you can do with no-code tools now, even like build simple applications. Level two is going to be intermediate, will involve some code, and level three, we're going to be building full-on applications as well as implementing techniques like rack. We'll be using lots of different models like your open-air models, CLA, as well as open-source models like Llama, different APIs and tools like Lang chain. By the way, in case you don't know me, hello, my name is Tina. I used to be a data scientist at Meta. Now I create content, and I run a program called Lonely Octopus, where we teach people AI and data skills to work on real freelance projects. Also, in case you're wondering, the reason why I have this on is because I ran into a pole and

fractured my hand all right. Without further ado, let's get started. A portion of this video is sponsored by HubSpot AI personal finance chatbot. Recently, I got really obsessed with this anime called 886. I watched the entire anime in like 2 days. Then I had to buy all the light novels, and I had to buy all the figurines, all the posters, like everything. Anyways, amazing anime. It will make you cry. Go check it out if you're interested in anime. But I realized that my spending habit is out of control, and I really need to come up with a new financial plan and like a new budget and things. So I'm in the market for a personal finance AI tool, like a chatbot that can help me get my finances back on track. Basic skills you will need are prompt engineering, and you can use it on ChatGPT or a platform like GPT-4, or you can also use open-source models. Additional skills for level two and three are Python, AMA, some sort of front-end dashboard like Plotly, and if

you're feeling extra challenge, and you want to build your own custom web app Lang chain, mostly for rag, which is retrieval augmented generation and flask. Level one is just prompt engineering. So first, you need to go to your bank account and download these statements within the period of time that you're interested in for the prompt. I like to loosely follow the 5 W framework, which is who, what, when, where, why. So for the who here, you are a personal finance analyst and adviser with the personality of my guy from Naruto, someone that is very positive and is perceptive of the feelings and goals of other people. All right, this time you'll be picked. I just know it. You ready, Lee? I need some positive vibes when confronting my spending habits. For the what, it's your goals to help me better understand my spending habits to achieve my financial goals of buying a house by the end of 2025. So it's kind of also the why, which is buying a house, and it's also when, um, by the end of 2025. And for the where, where

I'm currently at is I am missing $200,000, but I'm struggling to come up with a plan that I can stick to. I have attached my financial statements, so I'm also going to attach my financial statements. All right, dynamic entry, youthful determination, let's dive into your transactions like a leaf ninja, so motivational analysis. Your financial journeys just like riger's training, consistency and precision are key. We need to analyze where your hard ear cash is flowing and find areas where we can tighten the belt to save for that dream home by the end of 2025. So, steps to take is to examine the income sources, scree expenses, creating a budget, and tracking progress makes sense um. So, the analysis is income and expense overview. So, looking at the income and the expenses and then categorize the expenses, essentials and non-essentials, and identify saving potentials, uh cut down non-essential spending and optimize essential spending. Let's first summarize your total income and total expenses, so total income and total expenses as well.

as the ending balance. By the way, if the numbers are looking really big, that's because this is also including business income and business expenses as well. So yes, at least our balance is positive. Okay, okay. So then it goes on to categorize the analysis. For some reason, it categorized everything as non-essential. So what we're going to do is say that not all of it is non-essential. Maybe we can start by roughly categorizing things and then deciding what is essential and non-essential. List out the categories. Okay, so key expense descriptions: American Express payments, DoorDash, rent, utilities, grocery store, subscriptions, medical, loan payments, and transport. So for the American Express payments, what I will do is I will go to my AEX and then download everything from there as well to Mac. What is here from my Bank of America statements? So this works fine, but maybe you don't want OpenAI to be collecting your financial data. This is where you can consider using something like GPT for all. You can use it similarly to chat.

GPT, but you can download different types of models, including open-source models like Llama 3.1. Add in your finance statements here and chat with it. Level two, you're going to have some code. First, you want to download AMA, which will allow you to programmatically download open-source models like Llama 3 and allow you to use them locally on consumer hardware. It's kind of similar to GPT for all, except you're able to directly access these models through code. Now you can do things like loop through the statements, do some summarizations, and group things by categories. Okay, so so far you might be like, Yeah, but this is basically what we were doing at level one. Well, yes, that is correct, but by doing it programmatically, the first benefit is that we have a lot more control. You can do things like clean up the data a little bit, you can combine different statements together easily when you're looping through it, you can also specify what are the transaction types that you want to do. Secondly, and more importantly, it allows you to take that

data and then create something else out of it. For example, a dashboard. Level three is making your own custom web app with your own custom RAG. RAG, which is Retrieval-Augmented Generation, is a technique for enhancing the accuracy of AI models. It allows a model to access a specific data source, so when it's talking with you, it'll be referencing back to the information that's stored, as opposed to just trying to come up with the answer based upon the original data it was trained on. Platforms like ChatGPT and GPT-4 actually have that built-in already, so that's why you're able to upload your statements and then interact with it based upon the statements there. But since we want to implement our own custom web app, which, by the way, you might want to do if you're actually trying to build a consumer-grade product like a personal AI finance bot that you can actually sell, you need to implement your own RAG. LangChain is great; it offers you a suite of different tools in order to implement RAG. Then you can use something

like Flask to put everything together and make your custom web app. Something I'm sure you've heard me talk about before, if you've ever watched me before, is how to leverage AI tools like ChatGPT to help with both your professional and personal tasks. I found it especially relevant to increasing my productivity, something that's a crucial skill in keeping up with the future. HubSpot has a whole bundle of resources designed to help you understand and make the most out of ChatGPT. It will make you super productive, it's super well laid out, and gives you ideas on a lot of ways you can incorporate ChatGPT. For example, using ChatGPT to consolidate research materials like different articles, tutorials, all types of resources. You can directly ask it questions instead of having to flip through all of these different resources yourself or keeping up to date with industry news and updates. They also have this nice flowchart on when it's appropriate to use ChatGPT to solve a problem or streamline your workflow. My favorite part is that it gives you over 100

prompts to get started with ChatGPT, which is great since prompts are so important to using AI to the fullest. I highly recommend that you download it using this link over here, also linked in the description. Thank you so much, HubSpot, for providing free resources to help us leverage the power of AI and for sponsoring this portion of the video. Now, back to the video. Your personal AI assistant with video. So, it's already very cool that you can use something like ChatGPT to talk to it and ask it questions through text. You can also use audio functions, and you can even send it images. But sometimes, you just really want to communicate using video. Maybe you're on a first date, and you have poor social cues, and you're not really sure what your date is feeling like, what's the right thing to say. So, maybe you can take a video of her or him and be like, What are they feeling right now? What should I say? How's everything going? Not creepy, actually. That's pretty

creepy. Never mind. Okay, other use case. Say like you're wanting to learn ASL, like American Sign Language, and you want to be able to show what it is that you're doing and see if you're doing it correctly. And speaking of learning languages, if you are an elder person like me and you are trying to communicate with the youths these days, the Gen Alphas, um, you may find that you have no idea what the they're talking about. Mad lit, on God, on God, no cap. I have no idea what's going on here. So wouldn't it be nice if I could take this and give it to AI and be like, Can you explain to me what is happening here? So basic skills is prompt engineering. Additional skills include Python, the OpenAI API, using GPT-4, Whisper, and text-to-speech. And if you're feeling extra fancy, Shiny, which is a package for building interactive web apps, and you can host your assistant on the internet. So level one is just prompt engineering. At the time of this video, um.

Open AI did an announcement like a couple of months back, saying that they're going to actually release this functionality, this video functionality, where you should be able to communicate via video in real time. Well, can you take a guess at what I might be doing based on what I'm showing you here? H, from what I can see, it looks like you're in some kind of recording or production setup. However, they announced like two weeks ago that they're actually going to delay this functionality, and we're not sure like when it is that we're actually going to get it. So, at this time, you can't directly upload a video um into GPT 40. What is being shown on the poster in this video? I can't directly view video content, but I can help it. Yes, you can analyze it frame by frame. There you go. I've extracted and displayed several frames from the video. If you can identify the poster from any of these frames or if you need a closer look at any specific frame, please let me know. Last frame? No, here's a closer look at

the describe the Poster's contents briefly. The poster in the last frame of the video seems to be challenging to analyze for text content using OCR, likely due to resolution or boo honestly. As you can see, not great. It kind of drops different frames, and it doesn't really know what's happening that well. And other um AI models, even worse. Most of them don't even allow you to upload videos, so we can build something similar by ourselves. And by the way, if by the time you watch this video, OpenAI has already released that model with the video, you may still want to consider building this product because you might not want OpenAI to collect your video data as well, or if you actually want to build your own application. So level two is how we're going to implement processing video information and the accompanying audio to access the OpenAI API. So you can't directly just put in the video to the API. You first need to take the video and just extract the audio part using Whisper One, and then you convert that to

text which we will send as text input to GPT-4. Then you want to take the video part and feed it in separately as frames of images. Then you use OpenAI text-to-speech in order to take the response and then speak it to the user. To put everything together, you also need to set a system prompt, which is the prompt that kind of governs the entire AI model, so it's able to make sure to split the video into different components and then put it all together. I will leave the system prompt in the description below if you want to check it out. All right, level three. So now I have something that functions. It's basically a recording of a video, which then you're passing it along to OpenAI's APIs and then getting it back as audio. However, it's kind of ugly because it's just like, you know, on this Python notebook. So that's why level three is when we make it pretty and actually make it a proper application. That's also when you use Shiny. Shiny is a web application framework, so you're

able to have this web app where you can take a video and record yourself, um, and then afterwards it would be able to speak directly to you. You can also easily put in different UI components and deploy it. Yay! Now you can directly interact with your AI assistant using video. What does this painting say? The video showcases a sign that emphasizes the key components of success. It lists hard work, persistence, late nights, rejections, sacrifices, discipline, and criticism at the top, highlighting these as essential elements. Then it also includes doubts, failure, and risks further down, showing that these challenges are part of the journey towards success too. It's a powerful visual reminder that success isn't just about the positive aspects but also about overcoming difficulties and setbacks. Personal AI travel app. The top thing on my bucket list right now is to drive from the west coast of North America all the way to the east coast. And I recently just took my driver's license test, um, and I should be getting my license soon. I'm honestly so excited, uh, especially because recently I've been

watching a lot of like Journey anime, for example, Fen. Let me know in the comments if you've watched Fen. She's like a thousand-year-old elf, and her and her companions travel across this fantasy landscape. There's also Kino's Journey, where Kino, just with her gun and her talking motorcycle called Hermes, travels across the world, learning about different people, different customs, the beauty, and also the bad parts. There's also Somali and the Forest Spirit. Okay, okay, I will stop talking. Anyways, I'm just so excited to take this road trip, and I'm starting to plan for it now, which is why I want to build a personalized AI travel app so that I can plan an itinerary that has all the places I want to go to, has the right budget, and the best restaurants, and try to make it to things like concerts and festivals as well. Basic skills for this project: just prompt engineering. You can use ChatGPT for all your favorite AI chat interface. Additional skills include Make.com, which is a no-code platform.

that allows you to automate different flows using different types of plugins. I was honestly really impressed by this platform considering that it's no code, and I would say you can build like maybe 80% of it just by using this tool. But of course, if you actually want to make a really good product and potentially even a commercial product, you will need Python, the OpenAI API, your choice of model, some sort of front-end dashboard like Streamlit, and other types of APIs in order to incorporate different sources of data. Level one prompt engineering roughly using our 5W framework again: who, at, when, where, why. You are a personal travel agent with the goal of designing a custom travel plan from West Coast Vancouver to the East Coast. I want to start Vancouver, Canada, because that's where my car is going to be licensed. Okay, I couldn't not add this part. Um, I've been really into anime like Fate/Zero, Kino's Journey, and Somali and the Forest Spirits, so I like to replicate some of their travels but on Earth, like deserts, caves, beaches, big libraries. Ask.

me additional questions until you have a good understanding of my interest, traveling preferences, timelines, budget, etc., to suggest a detailed plan for me. Here we go. We answer all these questions, and voila, a draft itinerary. Week one: Vancouver to the Rockies, including Stanley Park, local seafood, cozy cab, Whistler, hiking in the summer or skiing in the winter, Art Museum, Jasper, Rockies to the Prairies, hiking to B National Park, a Lakeside Cabin, a vibrant art scene, and a quaint bed across the Prairies, Saskatchewan, Manitoba, and all the way North, finally ending in Journey's End to the east coast, wrap the journey to St John's where Halifax. You can ask to change the format of the itinerary, ask it to elaborate on certain things, and then just interchange things in and out too. Level two, so this is great. We have this itinerary here within chat GPT, and we can look at it, but it'll be nice if I actually could have a way to document when the dates are and like

more details about these specific events. That is where make.com comes in. You can take these itinerary items, translate them into calendar events that you can put onto your Google calendar, and then also put these items on Notion. So this way, you can add more detailed information for all of these items and have them all nicely organized on Notion, so you can look through them and edit them easily. That's pretty cool, right? Honestly, really impressed for a no-code tool. But to make it even better, wouldn't it be great if you could visually see this and, as you're traveling, be able to track your journey? That is why, for level two, we're going to get into some code. You can use a Jupyter or Colab notebook and use your model of your choosing. If you want to use OpenAI, you can use OpenAI APIs, open-source models. You can use AIMA to download the models that we talked about earlier and then connect that to Streamlit, which is a really nice little Python module that allows you to create a dashboard. So this is sort of

like the foundation, and in this way, you can start showcasing stuff onto the dashboard. But what really makes it magical is level three. This is where you can start incorporating other sources of data. For example, if you want to track your journey, you can use the Google Maps API and be able to label where all of these things are and be able to click into them. You can also incorporate APIs like the weather, so you're able to know what the weather is like wherever it is that you're going to be staying. Information about certain events or festivals, you can even incorporate where the restaurants are, which ones are the best restaurants, the ratings of the restaurants in different places. Yeah, the world is your oyster. You can incorporate all the information that you care about and use that information to plan out where it is that you want to go, what's the order of doing things, and specific places that you want to visit. All of that can be showcased visually on this dashboard, like a truly custom AI travel app. AI music composer, if you watch Short

from content you might have heard this audio that went viral recently. How the hell you spell. Did you know that this is actually AI generated? I was pretty surprised when I first found out and after playing around with some of the AI music generators. I was like damn that is pretty good. As someone who's not particularly musically inclined. I honestly can't tell a difference between pop songs these days and these AI generated tracks. So for this project we're going to make an AI music composer. So I was thinking why not play around more generating AI tracks and potentially even like putting them together into playlists. This could actually be really useful like for example um I do the study with me live streams and I like to play lowii music when I'm studying. But I'm pretty Limited in terms of playlist choices because most things are copyrighted. So if I use AI generated soundtracks. I could make myself a playlist that is not copyrighted and I can customize it to however it is that I want. That is why so in this project we're going to make

playlists that are completely AI-generated that can be used for free for live streams, background music, or just like music to vibe to when you're studying. Basic skills, super simple. We're just going to use AI generation platforms like Udio and Sunno. Oh, and you also need decent taste in music. Additional skills, you know the drill: Python, the Udio API, and the OpenAI API to use GPT-4, text-to-speech, and DALL-E for image generation, and your choice of front, for example, Streamlit. Level one, we're just going to play around with these AI generation platforms, Sunno and Udio. Considered best in the industry, they're honestly pretty similar. I think Udio is slightly better, and also I like Udio more because they do have an API, which we'll be using for level two. But yes, level one, just try around, use different prompts, try out the different features, see how you can put in your own lyrics, generate different lyrics, set the mood for different things, see what are the controls that you can

have sh underneath your bed in the shadows. Okay, level two. So, okay, technically you can download all of the tracks and then put them together to make a playlist yourself, like using some sort of sound mixing thing, but you can also do it programmatically by using udo's API. This here is their API. They list out all the things that you can do. It's honestly pretty simplistic, but the one that we're interested in in particular is being able to generate links in order to share the musics. You can use Python to access their API and take the link of the music that you generated and practically put them together into a playlist. So, I'm not really a music person, so you probably have to tweak things around to make it actually sound good and flow well and things like that. You can do it programmatically, like with these libraries, or I guess you can just import it into your normal DJ software and put it together. Definitely a lot less tedious than downloading everything manually and putting them all together, like 50 songs. So, ideally

speaking, it would be so cool if we can also programmatically create music, like tweak different, uh, text and, like, different parameters, and then just try out a bunch of different ones and see which one works the best, right? Like, maybe loop through all the different settings and find the right combinations. Unfortunately, the audio API does not allow you to do that, and I couldn't find one that does. If you find something like that, please do leave a comment. I would love to check it out. But at least we have a playlist now. [Music] Yay! Now we can go to level three to take it to the next level. So, we have this playlist now, and you can tweak it, and you can make it relatively good, but what would make it even cooler is if we can generate accompanying imagery based upon the lyrics and the emotions that the music invokes. Something like that. You can even put it on YouTube as a playlist, like similar to the playlist that is currently going on in the background here, like you have some music.

that's playing. You guys can't hear the music right now because it's muted, but there is some music that is playing. You can have different imagery show up based upon what is happening in the lyrics, and you can change the style to be whatever it is that you want. It can be like anime style, it can be realistic, it can be Loi style, whatever you want to do this. You take what you did from level two, and then you put it through Whisper from OpenAI in order to convert it into text. Then you can take that text and put it into GPT-4 in order to get the imagery and the mood of the different songs. Then you can ask the model to create a prompt for DALL-E to generate images related to the mood and imagery of the songs. There you go. Finally, you just need to write a script that puts together the audio and the visual so that you have a full playlist. [Music] Final project workbooks for books, book workbooks. I got inspiration for this project when I was rereading one of my

favorite personal finance and business books called the millionaire fast lane. Highly recommend if you're interested in personal finances, becoming financially free, all those things. Anyways, this is a really dense book, and there's just so much good information, and there's so many exercises that the author MJ tells you to do throughout the book as well. So in my copy, I have these little scribbles where I write down the answer to some of these and like some thoughts and stuff. But I thought to myself, wouldn't it be nice if there was like a workbook that actually compiled together all the different questions and exercises that he had so I can just fill them out there instead of, you know, trying to figure out where it is I scribbled random things. So then I thought, oh, maybe there is, and I looked, and I googled mire fast lane workbook, and I found one on Amazon that was selling for $1.99. Unfortunately, it only had one star, so I figured it probably wasn't really good. Um, but that kind of gave me

the idea like AI would be really good at generating these workbooks for people. You can definitely use it for personal use. I don't know about commercial use. Not a lawyer, not lawyer advice, not Financial advice. I don't know. So in this project, we're going to make AI-generated workbooks. How I would approach this level one just prompt engineering. So first, you got to find a PDF version of the book you're looking for, um, or like, you know, images, whatever it is, some sort of version. You might be able to find some of the PDFs online, or if you just have a version of the book, you could like take pictures of it. I don't know. But technically, you can find a lot of these PDFs online. Assuming that you found the PDF in a fashion that is legal, all you have to do is feed it to your favorite large language model and create a prompt in order to summarize the book and ask it to extract the questions that are being asked in the book and put it together into a workbook that you can

then download as a PDF. In this particular case, this book is over 300 pages, so it's a relatively mid-length book. I couldn't manage to feed it into Chat GPT using GPT-4 entirely, so I had to actually split it up into different chunks and then feed it in and ask it to summarize, create a workbook, and then put it all together again. It's a little annoying, but it does work fine. But if you want to be able to just do it in one go and not have to go through that process, you can also check out other models that have larger context windows. For reference, the standard GPT-4 has a 128,000 context window, and they probably CLA, on the other hand, does have more at 200,000. But I believe the model that has the biggest context window is actually Gemini by Google. It says that it has up to 1 million tokens. Or if you can find any open-source models that have bigger context windows that can do this, please link in the comments. I would love to check

it out because this is such a useful use case for me. Level two, let's do this programmatically. So the strategy we talked about earlier, I think it's fine for most use cases, um, but for some reason, if you want to program automatically, like mass-produce a lot of these different workbooks for different books, level two is for you to do this. We're actually going to use another no-code/low-code tool called Bubble. Bubble is a full-stack no-code app builder that connects with a lot of different APIs. So it's actually pretty similar to Make.com, but it focuses on building web apps, not automations. So yeah, you can use Bubble to connect to the OpenAI API, put in the prompt that we talked about earlier, then take that data and use it to populate the web app. Bubble's pretty cool because not only can you do the backend stuff like connecting to OpenAI, um, it also has really cool UI features that you can just drag and drop to make the front end look really nice as well. You can even

have signups and logins now finally level three. So in this level, you want to go beyond what bubble is capable of. Maybe you want more flexibility in the way that things look. Maybe you want to be able to edit the workbooks programmatically as well. That is why you're going to use the API from pdf. Co. Pdf. C is a low code and rest API that allows you to do stuff like converting PDF, editing, extracting, and automating it. So after getting the workbooks from open AI, you can directly edit the PDF so that you can make it look like the way that you want. And you can actually reformat it as HTML templates. You can make them even to be editable and save the responses in the work buts um in a database. For more details and resources, check in the description. All right, that is all I have for you guys today. Thank you so much for watching. I hope you enjoyed these five different projects and link your GitHub links if you end up doing any of them. I would love to check.

them out. Also, put other project ideas that you have too. I will see you guys in the next video or live stream.

5 Unique AI Projects (beginner to intermediate) | Python, LangChain, RAG, OpenAI, ChatGPT, ChatBot

Hi! My name is Tina and I'm a ex-Meta data scientist. Now I create content and other internet things! 

This channel is about coding, tech, career, and self-study. I love learning new things and my goal is to always minimize effort and maximize outcome! 

I strive to do things in the most efficient way possible and with the highest reward. I basically refuse to do extra work to achieve my goals c: 




5 Unique AI Projects (beginner to intermediate) | Python, LangChain, RAG, OpenAI, ChatGPT, ChatBot

5 Unique AI Projects (beginner to intermediate) | Python, LangChain, RAG, OpenAI, ChatGPT, ChatBot

스크립트