About this video

• Video Title: The ONLY AI Tech Stack You Need in 2026
• Channel: Cole Medin
• Speakers: Cole Medin
• Duration: 00:34:25

Overview

This video provides a comprehensive overview of Cole Medin's AI-first tech stack, detailing the tools and technologies he uses for building software, particularly AI agents. He explains his reasoning for choosing specific tools, offers alternatives, and categorizes them by their application, including core infrastructure, AI agents, RAG agents, web automation, full-stack development, and deployment.

Key takeaways

• AI-First Approach: The speaker emphasizes an "AI-first" philosophy for building software, believing it's the most effective approach in the current landscape.
• Core Infrastructure: Essential tools like PostgreSQL (with Neon/Superbase), Redis (and Valke for caching), and Claude Code (for AI coding assistance) form the foundation of his stack.
• Agent Development: Py Pydantic AI is highlighted as the primary framework for building individual AI agents, while LangGraph is used for orchestrating multi-agent workflows. Arcade is crucial for agent authorization and tool security.
• RAG and Data Handling: For Retrieval Augmented Generation (RAG), Dockling and Crawl4AI are used for data extraction, PostgreSQL with PGVector for storage, and Memzero for long-term memory. Neo4j and Graffiti are recommended for knowledge graphs, and Regos for RAG evaluation.
• Web Automation and Full-Stack: Crawl4AI and Playwright are key for web automation, while Fast API, React, and Shad CN/Tailwind CSS form the basis of his full-stack development. Streamlit is used for rapid prototyping.
• Deployment and Infrastructure: Render is preferred for its simplicity in deployment, while RunPod is recommended for GPU-intensive workloads, and DigitalOcean for managing virtual machines. Docker is the standard for containerization, and GitHub Actions for CI/CD.
• Emphasis on Problem-Solving: The speaker repeatedly stresses that the ultimate goal is to solve problems, and the choice of tools should support that, rather than becoming an end in itself.

The speaker utilizes PostgreSQL in several ways:

• Core Database: It's his primary database for most applications, often hosted through platforms like Neon or Superbase.
• Vector Database for RAG: Thanks to the PGVector extension, he uses PostgreSQL as a vector database for RAG agents. While acknowledging that dedicated vector databases might be faster, he prefers PostgreSQL because many RAG strategies also require a regular SQL database to store document or user data.
• Long-Term Memory Integration: He uses Memzero for long-term memory, which integrates with PostgreSQL (specifically PGVector), allowing memories to be easily added, searched, and injected into system prompts.
• Full-Stack Development: PostgreSQL remains his go-to database for full-stack development, including applications that support AI agents.

The speaker recommends the following tools for web automation:

For data extraction (pulling documentation ahead of time to store in a knowledge base):

• Crawl4AI: Described as fast, efficient, with LLM integrations, and capable of automatically cleaning junk data. He notes it's open-source and feature-rich.
• Dockling: Recommended for more complex documents like PDFs with diagrams and Excel files.

For live web interaction (giving AI agents tools to pull information from the web in real-time):

• Crawl4AI: Can also be used here for agents to extract text from a specific URL live.
• Firecrawl: Mentioned as another good alternative for live web interaction.
• Ampify or Bright Data: Recommended for social platforms like LinkedIn, X, and Instagram, which are not handled as well by other tools.
• Playwright: Used for simpler browser automations, web testing, and visual validation by AI coding assistants. He considers it the "deterministic web automation king" and highlights its Playwright MCP server.
• Puppeteer and Selenium: Mentioned as other good options for browser automation, though he has largely switched to Playwright.
• Browserbase: Used when an agent needs to control a browser live for complex tasks. It offers managed infrastructure, anti-bot detection, and records sessions for review. Stagehand and Director are tools built on top of Playwright within Browserbase for AI-driven browser control.

The key distinction is the timing and purpose of the web interaction:

• Data Extraction: This is a pre-processing step to gather information that will be stored and later used by agents (e.g., for RAG).
• Live Web Interaction: This involves the AI agent actively using a tool to navigate the web and retrieve information on demand as part of its task execution.

The speaker outlines several key considerations for deployment and infrastructure:

Deployment Platforms:

• Render: Recommended for its simplicity and ease of use, especially with its YAML-based infrastructure-as-code and Git-based deployments. It's good for frontends, background workers, and cron jobs.
• Fly.io, Netlify: Mentioned as other Platform-as-a-Service (PaaS) options.
• AWS or Google Cloud (GCP): Recommended for more enterprise-level requirements, SLAs, or compliance needs. GCP is specifically mentioned for its serverless functions.

GPU Workloads:

• RunPod: Highly recommended for reliable and instantly available GPU hosting at a competitive price. Offers spot instances for lower costs.
• Tensor Dock: An alternative for very cheap but potentially less reliable GPU hosting.
• Lambda Labs: Another alternative for GPU hosting.

Virtual Machines (VMs) / Self-Managed Infrastructure:

• DigitalOcean: Used for owning and managing the machine, particularly for hosting the "Local AI" package in the cloud. Praised for reliability and predictable pricing, with growing AI integrations.
• Hostinger (KVM offering) or Hetzner: More affordable alternatives to DigitalOcean for VM hosting.

Containerization:

• Docker: Considered the industry standard for deploying applications, creating isolated environments, and solving the "works on my machine" problem.
• Podman: A good alternative to Docker, especially concerning licensing costs, though it may be less robust.

CI/CD:

• GitHub Actions: The preferred tool due to its integration within GitHub repositories, free tier for public repos, generous pricing for private repos, and a large marketplace of actions. AI coding assistants can help generate the necessary YAML.

Testing:

• Pytest (Python): Standard for Python testing.
• Jest (TypeScript): Standard for TypeScript testing.

AI Code Review:

• Code Rabbit: Highly recommended, especially for open-source projects due to its free, detailed, and automated code reviews, including security vulnerability detection.