AI tooling moves fast, and most of it is noise. These are the categories — and the tools inside them — that have earned a permanent place in how our team designs, builds and ships AI-powered software.

1. Coding assistants

Agentic coding assistants such as Claude Code, Cursor and GitHub Copilot now do far more than autocomplete. They read a codebase, run commands, fix failing tests and explain unfamiliar code. The win isn't writing more code — it's spending less time on the boring parts and more on architecture and review.

  • Give them context: a short project README with conventions pays for itself daily.
  • Review everything. Treat suggestions like a pull request from a fast, new teammate.

2. Model APIs

We stay model-agnostic and pick per task: frontier models such as Claude, GPT and Gemini for reasoning-heavy work, and smaller or open-weight models for high-volume, narrow jobs where cost and latency matter. Keeping a thin abstraction over providers means switching is a config change, not a rewrite.

3. Orchestration

LangChain and LlamaIndex are useful for wiring retrieval, tools and memory together quickly. For production we keep the orchestration layer thin and explicit, so every prompt, tool call and fallback is easy to read and test.

4. Vector databases

For retrieval we reach for Pinecone or Qdrant when we need a managed, scalable index, Weaviate when hybrid search is central, and Postgres with pgvector when the data already lives in Postgres and the scale is moderate. Simpler infrastructure usually wins.

5. Evaluation and observability

This is the category most teams skip and most regret skipping.

  • Evals: a versioned set of real inputs with expected outcomes, run on every prompt or model change.
  • Tracing: tools like Langfuse or LangSmith record every model call, its inputs, outputs, latency and cost, so you can debug a bad answer in minutes.

6. Model Context Protocol (MCP)

MCP gives AI assistants a standard way to connect to tools and data — databases, CMSs, design files, ticket systems. Instead of building a custom integration for every assistant, you expose a service once as an MCP server. We use it to let assistants query and edit content in our CMS directly.

7. Speech and media

OpenAI Whisper for transcription and ElevenLabs for natural voice cover most speech needs, from meeting notes to voice agents.

Tools change every quarter. The habits — clear context, small evaluations, good tracing — are what keep AI features reliable.

How to choose for your team

Pick one tool per category, wire up evaluation and tracing from day one, and revisit the stack every six months. If you'd like help choosing or building with these tools, our AI team is happy to talk.