Interactive guide · 2026 edition

AI/ML Tech Stacks, Visualised

A mind-map view of the 2026 AI/ML universe — LLM providers, frameworks, vector databases, RAG, agents, fine-tuning, eval, inference, and data. Pick your stage, filter by what matters, and see the stack that fits.

/ search 1-4 stage Esc reset

The landscape at a glance

— · click a node to jump to its card

Which stage are you at?

Pick one — the stack below adapts, and cards fade to only what fits.

All technologies

Decision cheat-sheet

Pick by use case

  • Chatbot: OpenAI / Claude + Vercel AI SDK. Streaming-first, great UX out of the box.
  • RAG app: LlamaIndex + Pinecone + Claude. Best data connectors and retrieval quality.
  • AI agents: LangGraph + Claude MCP. Stateful multi-step workflows with tool access.
  • Code assistant: Claude + fine-tuned model. Best-in-class code generation and analysis.
  • Search: Cohere Embed + Qdrant. Top embeddings with high-performance vector search.
  • Content generation: GPT-4o + Guardrails AI. Quality output with safety rails.

Pick by constraint

  • Lowest cost: Llama 3 + Ollama + pgvector. Open models, local inference, no API bills.
  • Fastest inference: Groq or vLLM. Custom hardware or optimized serving engine.
  • Best quality: Claude / GPT-4o. Top closed models for reasoning and generation.
  • Privacy-first: Self-hosted Llama + Qdrant. No data leaves your infrastructure.
  • Enterprise compliance: Azure OpenAI + Guardrails. SOC2, data residency, safety rails.

Anti-patterns to avoid

  • Building a custom vector DB instead of using Pinecone/Qdrant/pgvector.
  • Fine-tuning before trying prompting and RAG first.
  • Using agents for simple tasks that a single LLM call can handle.
  • Skipping eval — "vibes-based" quality assessment doesn't scale.
  • Treating all LLMs as interchangeable — they have different strengths.
Tweet Reddit LinkedIn Star on GitHub