Abstract:

This video provides a comprehensive overview of the creator's personal "AI-first" technology stack for software development, presented as a guide for 2026. Emphasizing a philosophy of "capabilities over tools," the speaker details his preferred technologies across several key domains. The stack covers core infrastructure like PostgreSQL and Redis; frameworks for AI agent development including Pydantic AI for single agents and LangGraph for multi-agent systems; and essential services like Arcade for authorization and Langfuse for observability. Specialized stacks are presented for Retrieval-Augmented Generation (RAG), featuring tools like Docling for data extraction and Neo4j for knowledge graphs, and for web automation using Playwright and the AI-driven Browserbase. The video also details a full-stack development setup with FastAPI and React, and concludes with deployment practices using Docker, Render, and self-hosting solutions like Ollama.

A Comprehensive AI-First Tech Stack for Modern Development

• 0:00:44 The Core Philosophy: The speaker advocates for choosing a stable set of tools and sticking with them, prioritizing problem-solving ("capabilities over tools") over constantly changing technologies.
• 0:02:28 Core Infrastructure: The foundation of the stack includes PostgreSQL (via Supabase or Neon) as the primary database, Redis for high-speed caching, Claude Code as the main AI coding assistant, and N8N for rapidly prototyping agent workflows.
• 0:06:41 General AI Agent Toolkit:
  • Agent Framework: Pydantic AI is used for building individual agents due to its flexibility and minimal abstraction.
  • Multi-Agent Systems: LangGraph is the tool of choice for orchestrating complex, multi-agent workflows, managing state, and enabling human-in-the-loop processes.
  • Authorization & Security: Arcade is essential for handling agent authorization, managing OAuth flows to securely grant agents access to user accounts on platforms like Gmail or Slack.
  • Observability: Langfuse is a critical tool for monitoring agents in production, allowing for tracking of costs, latency, tool calls, and overall performance.
• 0:12:08 RAG (Retrieval-Augmented Generation) Stack:
  • Data Extraction: The stack uses Docling for complex files (PDFs, Excel) and Crawl4AI for websites.
  • Vector Storage: PostgreSQL with the PGVector extension is preferred to combine structured data and vector embeddings in a single database.
  • Knowledge Graphs: Neo4j serves as the graph database, with the Graffiti library used for AI-powered entity and relationship extraction from raw text.
• 0:18:45 Web Automation Agents:
  • Deterministic Automation: Playwright is the primary tool for scripted browser automation and web testing.
  • AI Browser Control: Browserbase provides an advanced, managed platform for agents to control a browser live, handling anti-bot measures and recording sessions for review.
• 0:22:00 Full-Stack Development:
  • Backend: FastAPI is used for building APIs in Python, aligning with the language used for AI development.
  • Frontend: React and Vite are chosen for creating fast, lightweight user interfaces.
  • UI Prototyping: Streamlit allows for the rapid creation of simple UIs directly in Python, ideal for testing agents before building a full frontend.
• 0:27:04 Deployment, Infrastructure, and CI/CD:
  • Containerization: Docker is the standard for creating isolated and portable application environments.
  • Hosting: The speaker uses Render for simple PaaS deployments, Digital Ocean for virtual machines, and RunPod for affordable GPU-intensive workloads.
  • Automation: GitHub Actions is used to automate CI/CD pipelines, while Code Rabbit provides thorough, AI-powered code reviews for free on open-source projects.
• 0:32:11 Self-Hosting and Local AI: For running AI locally, the stack includes Ollama for serving large language models, Open WebUI for a local ChatGPT-like interface, and Caddy for simple HTTPS management.