LLM Engineering, RAG, & AI Agents Masterclass [2026]

Build a character AI chatbot with the OpenAI API, configure Anaconda and API keys, and guide the model with system prompts to adopt varied personas and explore tokens.

Install the Anaconda distribution, open a JupyterLab workspace, and configure the OpenAI API key via an environment file to build a character AI chatbot.

Send your first chat with the OpenAI API by configuring a GPT-4.0 mini model, building messages with roles and content, and reading the AI response in code.

Experiment with the OpenAI API to generate text, test different models from GPT 4.0 mini to GPT 4.0, and compare responses by asking factual questions or playful explanations.

Experiment with the OpenAI API for text generation by replacing messages and switching from GPT 4.0 mini to GPT four, then attempt problems first and compare with the provided solution.

Explore the OpenAI API response structure and token usage, including model details, response IDs, roles, and prompt and completion tokens.

Analyze two prompts with the OpenAI tokenizer tool to compare token counts for supervised vs unsupervised learning, then reflect on token efficiency.

Explore how the OpenAI tokenizer tool counts tokens, compares tokenization of examples, and shows how politeness like please and thank you increases token counts in a live notebook.

Set a system message to give the AI chatbot a persona, using a Python dictionary of characters like Sherlock Holmes, Tony Stark, and Yoda to shape responses.

Explore a final practice opportunity to swap AI personalities with the OpenAI API, updating a Python dictionary from Tony Stark to Yoda, then test and add a playful new persona.

Explore changing AI personalities via a practice opportunity solution by building a final Python dictionary entry, setting a sleepy cat system prompt for GPT-4.0, and testing with chat completions.

Explore how generative AI creates text, images, and code with the OpenAI API. Tailor AI personality with system messages and model choices, and learn tokens and prompts.

Build an artificial intelligence powered calorie tracker that identifies food from images. Estimate grams and calories and reveal macronutrient breakdown; use the OpenAI API with image input and base64 conversion.

Read and display a sample image using Pillow (PIL), inspect its format, size, and RGB mode, and prepare it for analysis with the OpenAI API.

Practice opportunity guides you to download a food image, save it in your project, update the image_filename variable, and load, display, and verify image format, size, and mode with PIL.

Learn to read and view images with PIL by loading a sample pizza slice image, checking its format, size, and mode, and displaying and saving a copy for analysis.

Master prompt engineering fundamentals by designing and optimizing text inputs for generative AI, covering context, instruction, input data, output indicators, and techniques like zero-shot, few-shot, and chain-of-thought prompting.

Explore prompt engineering fundamentals by drafting a prompt to classify a CEO's tone as optimistic or concerning from Solid Power Inc. earnings call transcripts, highlighting top phrases and actionable recommendations.

Master prompt engineering fundamentals by crafting context-driven prompts with input-output indicators, testing across AI models, and delivering tone assessments, evidence, and actionable recommendations for investment scenarios.

Encode the image to base64 and send it to OpenAI's vision model, then prompt what you see in the image using a GPT-4 default.

Draft a prompt for the OpenAI vision API to recognize food in an image for a calorie-tracking app, identifying salmon with asparagus, tomatoes, lemon, and dill.

Modify the food recognition prompt variable to ask a different question, run the cell again, and perform a sanity check to verify OpenAI API or vision GPT model responses.

Solve practice opportunities by tweaking prompts to call OpenAI vision models for analyzing a food image for calorie tracking, determining savory vs sweet, and listing colors.

Analyze food images with vision GPT models to estimate calories and nutrition per serving or per 100g, outputting a structured JSON with calories, fat grams, protein grams, and confidence.

Expand the structured nutrition prompt to include sugar grams and fiber content, test with pizza slices, and assess OpenAI's API nutrition estimates and confidence on complex dishes.

Expand the api payload to include sugar grams and fiber grams within the structured nutrition prompt, and test with a pizza slice and greek salad using OpenAI vision.

Review building AI-powered apps that process images and estimate calories. Master OpenAI's vision, prompt engineering basics, and defining Python functions for tasks.

Develop an adaptive AI tutor with a web interface using gradio, featuring a complexity slider to tailor explanations from simple to expert levels and streaming responses for real‑time answers.

Demonstrate how to set up gradio, the fastest way to demo an AI model with a friendly web interface. Explore pip install, instant sharing, and features like maps and streaming.

Build a basic artificial intelligence tutor function in Python using the OpenAI API and the GPT-4 mini model, without gradio.

Engage in a practice opportunity to test the AI tutor function with many personalities by changing topics—from gravity to the Titanic plot—and experimenting with tutoring styles and tone.

Practice opportunity solutions guide learners to test a modular ai tutor with multiple personalities, switch topics like gravity and photosynthesis, and review how questions are sent and answered.

Build a no streaming Gradio interface to wrap a Python function that acts as an AI tutor, with text box inputs and outputs, deployable to Hugging Face.

Explore how to configure Gradio interface components by adjusting the gradient interface title, description, and input lines, then run and restart the interface to observe changes.

Configure Gradio interface components for an AI tutor by adjusting lines, placeholders, labels, and titles, test functionality, and set up for streaming and sliders.

enable streaming in Gradio to show OpenAI API responses as they come, using a generator to yield each text chunk in real time for an enhanced chat experience.

Explore a practice opportunity to compare non-streaming and streaming interfaces, analyze a multi-paragraph question, and assess the delay differences in chat experiences.

Explore two radio interfaces, one with streaming and one without, and see how streaming delivers immediate multi-paragraph OpenAI outputs, enhancing user experience.

Learn to build a multi-level AI tutor in Gradio by adding an explanation level slider that maps 1–5 to audiences, while streaming responses via the OpenAI API.

Explore testing ai tutor explanation levels from 1 to 5, then add level six, the PhD level Einstein mad scientist mode, and relaunch the Gradio interface to observe changes.

Test AI tutor explanations across six levels, from basic to Einstein mode, using a Gradio interface to adjust explanation complexity.

Develop a powerful streaming ai tutor interface using gradient-based ui and Gradio components, enabling interactive text boxes, sliders, and maps with an explain-level slider from project intro to deployment.

Compare frontier llms from OpenAI, Google Gemini, and Claude to build ai-generated websites and evaluate their math, emails, and landing pages using leaderboards.

Compare large language models by latency, context window, and benchmarks with the Vellum Leaderboard, exploring open source options like Llama and Hugging Face, versus closed source models like GPT-4.

Practice a hands-on benchmarking task on the vellum leaderboard by comparing Google's Gemini 2.5 Pro and Deep Seek R1 across latency, context window cutoff date, and accuracy.

Compare Gemini 2.5 Pro and Deep Seek R1 on a Vellum leaderboard, analyzing context window, latency, cost, and benchmarks to inform model selection.

Perform a blind test on Chatbot Arena to benchmark AI models through crowdsourced voting, comparing arena scores, cost, and performance in side-by-side prompts.

Act as an artificial intelligence judge to compare blind responses from two models in the chatbot arena, evaluating clarity, creativity, and correctness across technical, creative, and funny prompts.

Walk through the practice opportunity solution in the chat bot arena, testing prompts like a birthday poem for Layla and electric-car jokes across OpenAI and llama models.

Set up API keys for Google Gemini, OpenAI, and Claude, install libraries, configure API clients, and compare math and creative abilities across Claude, Gemini, and GPT.

Compare the code generation abilities of three LLMs—Sonnet by Anthropic GPT-4, OpenAI, and Gemini 2.0—by configuring the model and testing a BMI calculator function with metric and imperial units.

Compare LLM coding abilities by building a Python BMI calculator, testing OpenAI, Claude, and Gemini models with a structured prompt and docstrings, including unit handling, then compare latency and cost.

Define the startup idea Connect Genius, an AI-driven CRM that analyzes customer interactions, predicts needs, and automates personalized follow ups, then build a reusable HTML prompt for landing pages.

Practice structuring prompts by iterating startup ideas, updating an HTML prompt variable, and previewing generated web pages and landing pages without using models yet, preparing for later tasks.

Review the practice opportunity solutions and customize a startup concept: Quantum Leap Computing, a cloud platform offering affordable quantum computing resources for researchers, plus a user-friendly interface and educational materials.

Generate websites and html landing pages by sending an html prompt to the OpenAI api's chat completions, then save and review the resulting html page.

Experiment with the model parameter by switching to GPT-4.0 mini, rerun the cell to compare HTML output, and confirm the updated OpenAI landing page HTML after a browser refresh.

Demonstrates generating an HTML landing page for the practice opportunity by switching to the mini model, saving the file, and running code to produce a new OpenAI landing page.

Generate an HTML landing page using Google Gemini API's generate content method with an HTML prompt, then save to Gemini landing page HTML and review Gemini 2.5 flash output.

Compare raw HTML outputs from Gemini and OpenAI API using a chosen startup idea, updating your earlier prompt. Execute both APIs and report your findings.

Compare Gemini and OpenAI website generation by updating prompts and running AI-generated landing pages for a quantum leap computing platform, highlighting affordability, pricing, and AI-powered optimization.

Generate HTML landing pages with anthropic cloud using Claude 3.7 Sonnet, configuring the cloud client and prompts to produce and save the cloud landing page HTML.

Modify prompts to add a contact or sign-up section and test color cues to guide LLMs. Run OpenAI, Gemini, and Claude outputs to compare the cleanest semantic HTML structure.

Practice opportunity solution shows designing a website with LLM and a contact form. Explore using OpenAI API, Claude by Anthropic, and Gemini to update prompts and review saved prompts.

Compare frontier models like GPT, cloud, and Gemini with leaderboards and blind tests, then generate HTML landing pages using prompts and APIs across OpenAI Cloud, Gemini, and Anthropic.

Explore open source large language models on Hugging Face to build a chat-with-documents Q&A system using PDFs, Gradio, and models like Llama by Meta, Gemma Kwon, and Phi three.

Explore Hugging Face’s hub of pre-trained models, datasets, and spaces to tackle nlp tasks from text generation to sentiment analysis, with open-source options and offline use for privacy.

Explore hugging face for text generation, compare open-source models by parameters and popularity, and test the Phi3 4k instruct model on Colab by explaining Newton's second law to a five-year-old.

Explore Hugging Face text-generation models like Pi, llama, and Microsoft Phi three mini 4K instruct, most downloaded models, read summaries, test prompts, and explain Newton's second law to a five-year-old.

Install key libraries (transformers, accelerate, bits and bytes, torch, pdf, gradio), obtain a Hugging Face access token, and verify GPU availability for efficient 4-bit model deployment.

Explore Google Colab hardware accelerators by changing the runtime and compare CPU, T4, GPU, and A100 performance. Conduct external research to quantify speed gains from CPU to A100.

Compare CPU, T4 GPU, and A100 GPU on Google Colab to view speed gains of 5–10x and 20–50x over CPU. Note use cases, compatibility, and free vs pro availability.

Explore hugging face transformers pipelines for quick, beginner-friendly text tasks, and learn how pipelines, auto tokenizers, and auto model enable loading models, tokenization, and sentiment analysis with self-attention concepts.

Explore sentiment analysis in financial news using the Hugging Face transformers pipeline with a FinBERT model, analyzing two sample headlines and comparing results to the FinBERT inference API.

Apply the transformers pipeline to classify news as positive or neutral using a specified pipeline model, inspect sentiment scores, and test results with examples on Hugging Face.

Explore Hugging Face auto tokenizers in the transformers library and learn how to convert text into token IDs. Use from_pretrained to feed a tokenizer and compare tokens to OpenAI.

Experiment with auto tokenizer to load different pre-trained tokenizers (e.g., bert-based and another model), encode a sentence, print tokens and input IDs, and compare spacing differences in tokenization.

Compare Bert and GPT-2 tokenizers using the auto tokenizer to tokenize a sample text, observe differing token IDs and casing, and emphasize using the correct tokenizer to avoid garbage input.

Explore Hugging Face transformers auto model for causal LM to load pre-trained weights and run inference on GPT, llama, or gema models, using 4-bit quantization and tokenizer pipelines on Colab.

Experiment with different models and prompts by changing the prompt variable and testing questions like the capital of France or a cat poem, including llama 3.2 or 3 billion parameters.

Explore practice opportunity solution with transformers AutoModelForCasualLM by changing prompts and model IDs, running multiple models, loading tokenizers, and monitoring GPU memory to optimize performance.

Read pdf documents and extract text using the py pdf library, then join pages into a text string to enable questions and analysis of the pdf with the model.

Practice with the PyPDF library by printing the last 500 characters of a PDF, then compare them to the document and prepare a sanity check.

Explore the PyPDF library to extract and print the last 500 characters of a PDF, visualize the snippet, and perform a quick sanity check on the document end.

Build a q&a logic by loading an open source llm, extracting pdf text, and prompting with a structured template to answer questions grounded in the document.

Explore the practice opportunity to test the Q&A pipeline with an open-source LLM, exploring questions within a 6000-character limit and grounded in the document.

Test the q&a pipeline with an open-source llm by sending pdf text and a question, then extract the answer showing over 1.5 billion monthly users.

Build a Gradio-based pdf qa bot that switches among lightweight open-source large language models from the Hugging Face hub, using four-bit quantization and a model dropdown.

Visit the Hugging Face Hub, choose a small open-source model like Qwen 2.5, modify the code to add it to the available models, then test and comment on its performance.

Test the Qwen open-source LLM by selecting a model variant from Hugging Face and running it through a Gradio interface to validate with a sample question.

Students explore hugging face's role in democratizing ai with open source models, building a document based q&a app using gradio and transformers, and learning tokenization and quantization.

Classify news data and reason about sentiment using open-source Hugging Face models like deep seq, and apply a framework to select the right llm for your business with gradio interface.

Explore the Hugging Face datasets library, install key libraries like transformers and torch, and load financial news and traffic datasets to build a Gradio UI in Colab.

Explore the Hugging Face datasets hub, count total datasets, identify the most downloaded news dataset, and tally CNBC 2023 news datasets using the AI query in data studio.

Explore Hugging Face datasets, identify the most downloaded news datasets, and prepare a small training set for exploratory data analysis and visualization.

Load financial news data from Hugging Face using datasets, select title and description, concatenate into full text, and prepare for a deep sequence model in an inference-only workflow.

Examine the financial news dataset for extra columns like main domain and createdat. Adjust the display to ten samples, and access the full text of the 100th item.

Explore financial news datasets using the title and description with source, domain, and timestamp; visualize the first ten samples in pandas and access the 100th news full text for sentiment.

Load and test the 1.5 billion-parameter deep seek reasoning model, explore its chain-of-thought capabilities, and compare reinforcement learning with supervised fine-tuning for sentiment analysis using a Gradio interface.

Learn to load and test the DeepSeek reasoning model from Hugging Face, configure 4-bit quantization and auto device mapping, and build a text-generation pipeline with formatted prompts.

Test the math abilities of the DeepSeek model by solving simple and complex problems, such as counting apples after sharing, and explore how its reasoning differs from sentiment analysis.

Practice opportunity solution video demonstrates testing math capabilities of DeepSeek by solving a finance problem with the CAPM to identify a stock as undervalued.

Identify your business needs, pick the right model type and size, and choose open or closed source sourcing to build scalable AI with cost-aware inference and retrieval augmented generation.

Assess LM sourcing options— in-house, open-source LLMs, or proprietary cloud models—comparing open-source versus proprietary for cost, time to market, data provenance, and support to select the right model.

Compare open source large language models by performance using the Hugging Face leaderboards. Explore the shift from old benchmarks to new, harder benchmarks like drop, truthful QA, and if eval.

Compare old and new model benchmarks, explore parameter sizes, quantization, and pre-trained versus fine-tuned models, and review math, science, coding, and medical leaderboards for AI assessment.

Design a prompt and implement the analyze news sentiment function to produce financial reasoning and a sentiment classification for the full text (title plus description) with a topic tag.

Select a sample news item by index and analyze its sentiment with DeepSeek to establish a benchmark. Modify the test loop to analyze that item and verify the sentiment.

Practice opportunity solution demonstrates analyzing news sentiment with a DeepSeek pipeline, selecting a specific index, running sentiment analysis, and interpreting negative sentiment from peak gas season and oil pessimism.

Build a Gradio interface for a deep seq financial news analyzer, enabling fetch of random news, analysis, and display of the model reason and sentiment.

Explore deep seq models' structured reasoning, the role of hugging face datasets and leaderboards, and how clear prompts enhance explainability and output quality across AI projects.

Master retrieval augmented generation (RAG) pipelines in LangChain using 11 Madison Park data to ground answers with embeddings stored in chroma db and a gradio UI.

Master rag: link a retriever, orchestrator, and LLM to augment prompts with retrieved knowledge from external sources, delivering up-to-date, explainable outputs.

Learn LangChain 101 and key features for building multi-step LLM apps, chaining prompts, memory, tools, and RAG with agents and calculators to retrieve and analyze data.

Set up LangChain and tools, install Visual C++ 14, load data, configure OpenAI API, and build a rag pipeline with embeddings and chroma for 11 Madison Park.

Display the last 750 characters of loaded document and run a sanity check against the 11 Madison Park file using Python or GPT-4. Print the restaurant's email and phone numbers.

Demonstrate extracting the last 750 characters to verify the practice opportunity solution with LangChain textloader testing, visualize results, and perform a quick sanity check.

Split large documents into manageable chunks with the recursive character text splitter from LangChain, setting chunk size and overlap to improve embedding, storage, and retrieval in rag pipelines.

Explore configuring the recursive character text splitter by adjusting chunk size and overlap, rerunning cells, comparing chunk counts, and examining metadata for embedding and vectorization tasks.

Configure the recursive character text splitter to 500 chunk size and 150 overlap to reveal chunks and preserve context, and learn why source metadata enables citations in retrieval augmented generation.

Learn how embeddings convert words and documents into dense vectors. Create a vector store with OpenAI embeddings and search semantic meaning in a vector database.

Explore the TensorFlow embedding projector by identifying the nearest words to Italy and selecting two different words you like to visualize relationships.

Explore the TensorFlow embeddings projector by locating the nearest terms to Italy in a multidimensional space. See how PCA presents relationships and prepare to test retrieval in the next task.

Test the retrieval pipeline by querying a vector store and performing similarity search on restaurant document chunks. Validate that results reveal menu details, including plant based options and course menus.

Experiment with the retrieval pipeline by changing the test query variable and observing how the k parameter affects retrieved documents.

Practice opportunity solution demonstrates testing a retrieval pipeline with a vector store, adjusting k, and performing similarity searches to retrieve and display sources.

Build and test a retrieval augmented generation chain in LangChain by wiring a vector-store retriever with an OpenAI language model, then run a grounded query with sources.

Explore the chain's result dictionary, identify key information, then set temperature to 1.3 and test the or operator, and enable verbose and return source documents to reveal extra output.

Demonstrates a retrieval QA with sources chain, showing how changing temperature and enabling verbose and return source documents reveals top documents and their sources.

Create a Gradio interface for a rag pipeline that uses LangChain and the OpenAI API to query a QA chain, displaying the answer and its sources.

Expand Gradio interface questions to include bar reservations, dress code, and gift cards; customize the submit button text to Ask 11 Madison Park and re-test the clear all button.

Configure and test a gradio interface by adding questions, adjusting the title and submit, and validate the responses shown in the interface as part of the practice opportunity solution.

Create a retrieval augmented generation pipeline with LangChain, OpenAI embeddings, and chroma DB to deliver context-aware outputs via a Gradio interface, using a restaurant knowledge base like 11 Madison Park.