The AI_devs 3 course has provided an in-depth exploration of advanced concepts in artificial intelligence. It focuses on the integration and application of LLMs (Large Language Models) in real-world scenarios. This includes not only tools and techniques to enhance your understanding of AI development, but also data setup, working with LLMs, and building retrieval-augmented generation (RAG) systems. This article aims to summarize the key takeaways and insights gained from the course.

Disclaimer: this article is not affiliated in any way by the company behind AI_devs. I received neither compensation nor a discount. You won’t find any referral links here. At the time of writing this article, the course is not purchasable. I receive no benefits from advertising or promoting AI_devs, or its authors. Moreover, I paid the same price for the course as everyone else. The intention is purely to provide an unbiased, subjective opinion about the course.

How is the course organized?

The main training consists of five episodes, each containing five lessons. Every lesson appeared on a consecutive workday, so the main course lasted five weeks in total. There was also an optional, introductory pre-work week.

Almost every lesson ends with a task. You have to solve at least 80% of the tasks to get a certificate (here is mine). Note that these tasks involve communication via API. The API keys were separate for each participant. You have to write code (or instruct some LLM to write it for you) and execute it to pass. Code samples provided in the lessons are in JavaScript but you can use any programming language you are familiar with.

There were also extra tasks, for fun, that did not count towards the certificate. All the tasks (both normal and extra ones) used the CTF (Capture The Flag) form.

The most important concepts

Look at the course name suffix: agents.

That means it teaches how to use programming tools to solve problems automatically. Usually an agent uses many lower-level tools, such as some LLM via API and a database (not necessarily dedicated for AI).

It is important to handle unhappy scenarios gracefully. Giving up and showing an error to the user is better than presenting an incorrect or off-topic answer.

AI tools for development

There are a lot of tutorials and trainings on using LLMs and prompt engineering. In contrast, there is much less information about auxiliary utilities. These include powerful tools for monitoring and debugging AI applications, including specialized databases for vector storage, and utilities for web crawling. I’ll describe a few of them used in the course.

Read also: 10 Best AI Coding Assistant Tools in 2025 – Guide for Developers

FireCrawl: Web content extraction

FireCrawl is a web scraper designed for AI applications. It focuses on extracting clean, structured content from web pages. It can filter out noise like ads, navigation menus, and irrelevant elements. This makes it useful for feeding high-quality web content into LLM-powered applications. The tool can handle modern JavaScript-heavy websites and maintains proper content hierarchy. All of this makes it a powerful component for building AI agents that need to understand web content.

LangFuse: Monitoring and debugging AI applications

LangFuse is a monitoring and debugging platform for LLM-powered applications. It provides insights into token usage and costs. It can also analyze latency, and the performance of AI interactions. The platform allows debug prompts, and analyzes how they behave in production.

There are other alternative tools for prompt debugging and analyzing. For example:

  • LangSmith – Developed by LangChain, offering comprehensive debugging and monitoring features.
  • Portkey – Focuses on prompt management and optimization with A/B testing capabilities.
  • Parea – Provides analytics and monitoring with emphasis on prompt version control.
  • Helicone – Offers LLM monitoring with cost tracking and caching features.
  • Arize – An observability and evaluation platform for AI.

Vector databases

Vector databases are data storage systems designed to handle high-dimensional vectors. They are essential for applications involving machine learning and AI. They enable efficient similarity searches and retrieval of data. You can use them for tasks such as recommendation systems and semantic search.

Qdrant is a vector similarity search engine. It enables storing and searching through high-dimensional vectors using embeddings. The database offers filtering capabilities and real-time updates.

Speech-to-text tools

Speech-to-text (STT) technology helps applications to convert spoken words into written text. There are several tools which can be helpful in that matter:

  • Whisper by OpenAI offers transcription across many languages. You can run it locally or via API. 
  • AssemblyAI provides real-time transcription with advanced features like speaker diarization and content moderation. 
  • Deepgram specializes in real-time transcription optimized for specific industries and use cases. 
  • Happyscribe is another popular tool that offers transcription and subtitling services, providing an easy-to-use interface and API for seamless integration into various applications.

Text-to-speech tools

Text-to-speech (TTS) allows applications to convert written text into natural-sounding speech. This technology is essential for creating accessible applications and enhancing user experiences. There are tools for TTS as well:

  • OpenAI TTS provides high-quality voice synthesis with customizable options for tone and style.
  • ElevenLabs offers realistic voice generation with emotional intonation, making it suitable for storytelling and interactive applications.

Image generation

AI-powered image generation allows the production of high-quality visuals from textual descriptions or existing images.

ComfyUI is an intuitive user interface for interacting with various AI models related to art and image synthesis. It enables users to configure and run models without extensive programming knowledge.

Graph databases

Graph databases can efficiently store data structured as graphs, namely those consisting of nodes (entities) and edges (relationships). This structure makes graph databases suitable for applications that need deep connections between data points, such as social networks, recommendation systems, and knowledge graphs.

Neo4j is one of the most popular graph databases. It offers powerful querying capabilities through its Cypher query language.

Frameworks for agent creation

There are several frameworks which can help you create AI agents. For example CrewAI provides a straightforward interface for building agents that can interact with various APIs. 

  • Vercel AI SDK, likewise, is a powerful framework for building AI-powered user interfaces. It provides streaming responses and React/Svelte/Vue components, and has built-in support for popular AI models like OpenAI, Anthropic, and Hugging Face. The SDK makes it easy to implement features like chat interfaces with real-time streaming responses. It also offers type safety and handles rate limiting and error handling out of the box.
  • LangGraph focuses on integrating language models with graph databases, enabling developers to create agents that leverage complex relationships within data. 
  • Swarm (made by OpenAI, experimental at the time of writing) emphasizes collaborative agent behavior, allowing many agents to work together towards a common goal.
  • AutoGen offers tools for automating the generation of agent behaviors and interactions, streamlining the development process.

Code interpreters for AI

Code interpreters allow models to perform complex tasks, such as web scraping or data processing. Tools like BrowserBase provide a user-friendly interface for automating browser interactions, making it easier to gather information from the web.

Similarly, Playwright offers powerful capabilities for browser automation, enabling developers to write scripts that can navigate web pages, fill out forms, and extract data.

Interesting techniques

The development of AI applications requires specific approaches and methodologies. They differ from traditional software development. Here are some key techniques that have proven effective when building AI-powered systems.

Function calling

Function calling enables structured communication between the model and external tools or APIs. You provide a schema describing available functions and their parameters. An LLM can decide when to use specific tools and generate the appropriate arguments on its own.

This creates a standardized way for models to interact with external systems. For example, it can help searching databases, or controlling smart home devices.

One prompt for one problem

A key principle in AI development is to break down complex tasks into smaller prompts, rather than trying to achieve many objectives in a single go. Each prompt should be focused to address one specific problem or subtask.

This approach improves reliability and makes it easier to debug issues. For example, instead of asking an LLM to both analyze a document and generate a summary in one prompt, it’s better to split this into two steps. First analyzing the content, then creating the summary based on that analysis. This technique also reduces token usage which leads to lesser costs.

Wrap-up

The AI_devs 3 course has provided valuable insights into building AI-powered applications. From basic concepts to advanced agent implementations.

However, it’s important to understand that the field of AI development is evolving. New models, tools, and techniques emerge all the time. For instance, the Deepseek R1 model wasn’t even available during that course. Now, it is a notable player in the field. This highlights why continuous learning is essential for AI developers.

Taking one course, even an excellent one like AI_devs 3, is the beginning of the journey. You need to stay updated with the latest developments, constantly experiment with new tools, and refine your skills.

Based on my experience with AI_devs 3, I can recommend it, especially if you are interested in practical AI development. I’m looking forward to participating in AI_devs 4 when it becomes available.