Nishaglobal Education logo

The AI developer toolkit: a map

A beginner-friendly map of the tools used to build AI applications: models, model access, building blocks, agents, data, evaluation and operations. What each layer does, the popular tools in each, and how to decide whether you need a framework at all.

Lesson 39 of 4416 min readQuiz: 5 questions

🎯 By the end of this lesson you will be able to:

  • Explain what an AI framework is and why teams use them
  • Name the layers of an AI application and the popular tools in each
  • Decide when to use a provider SDK directly and when a framework helps
In this lesson
  1. What is a framework?
  2. The layers of an AI application
  3. Do you need a framework at all?
  4. How to choose between tools
  5. What this means for your role
  6. Key takeaways
  7. Quiz

What is a framework?

A framework is a ready-made set of code that handles common jobs so you don't have to build them from scratch. In AI applications, frameworks help you call models, connect them to your documents, give them tools, run multi-step workflows and check the quality of their answers.

You don't need to learn every tool in this module. The goal is to know what exists, what each one is for, and how to choose. The next lessons show each popular tool with a short, simple example.

The layers of an AI application

Quality & operationspromptfoo · Ragas · Langfuse · LangSmith · vLLMAgents & toolsLangGraph · CrewAI · Agent SDKs · Google ADK · MCPModel accessProvider SDKs · Bedrock · Microsoft Foundry · Vertex AI · OllamaModelsGPT · Claude · Gemini · Llama · Mistral · Qwen · DeepSeekApp frameworksLangChain · LlamaIndex · AI SDK · DSPyData & retrievalChroma · pgvector · Pinecone · QdrantYour application sits on top, using one tool (or none) from each layer
Most AI applications are built from the same layers. You choose one tool (or none) in each layer, depending on what you are building.
LayerWhat it doesPopular toolsCovered in
ModelsThe AI itself: understands and generates text, images, audio or codeGPT, Claude, Gemini, Llama, Mistral, Qwen, DeepSeekLesson 44
Model accessHow your code talks to a modelProvider SDKs (OpenAI, Anthropic, Google), cloud platforms (Amazon Bedrock, Microsoft Foundry, Google Vertex AI), local runners (Ollama)Lessons 41, 43
Classic machine learningTraining your own models on your own datascikit-learn, PyTorch, TensorFlow/Keras, Hugging Face TransformersLesson 40
Application frameworksPrompts, chains, retrieval and integrationsLangChain, LlamaIndex, Haystack, Semantic Kernel, DSPy, Vercel AI SDKLesson 41
Agents and toolsModels that plan, use tools and work in stepsLangGraph, CrewAI, Microsoft Agent Framework, OpenAI Agents SDK, Claude Agent SDK, Google ADK, MCPLesson 42
Data and retrievalStoring and searching your documents by meaningChroma, pgvector, Pinecone, Qdrant, Weaviate, ElasticsearchLesson 43
Quality and operationsTesting, monitoring and running AI in productionpromptfoo, Ragas, DeepEval, Langfuse, LangSmith, vLLMLesson 43

Do you need a framework at all?

Often, no. Every model provider offers a small official library, called an SDK (software development kit), that sends a prompt and returns an answer in a few lines of code. For many features, that plus your own code is all you need.

Start with the provider SDK when…Add a framework when…
You make one or a few model calls per requestYou need retrieval over many documents (RAG) with loaders, chunking and indexes
You want full control and few dependenciesYou want to swap between model providers easily
Your team is learning the basicsYou build multi-step agents with tools, memory and human approval
The feature is simple and stableYou need ready-made integrations with many data sources and tools

How to choose between tools

  • Language: most tools are Python-first; for web apps in JavaScript or TypeScript, check for official JS support (for example the Vercel AI SDK or LangChain.js).
  • Fit: pick the tool built for your main job, such as LlamaIndex for document Q&A or LangGraph for controlled agent workflows.
  • Maturity and community: active maintenance, documentation and many users make problems easier to solve.
  • Lock-in: prefer tools that work with several model providers.
  • Licence and cost: open-source, free tiers, and what managed services charge at your volume.
  • Your team and platform: a Microsoft .NET team may prefer Microsoft's tools; a team on Google Cloud may prefer Vertex AI and ADK.

A typical stack for a customer-support assistant

Model:        a mid-tier model from Anthropic, OpenAI or Google (via its SDK)
Retrieval:    LlamaIndex to load and index help articles, stored in pgvector (Postgres)
Workflow:     LangGraph: classify ticket → retrieve → draft reply → human approval
Tools:        an MCP server exposing order status from the company system
Evaluation:   promptfoo test cases run in CI on every prompt change
Monitoring:   Langfuse traces of every request, with cost and feedback

Each layer has one clear job. You could replace any single tool later without rebuilding everything.

What this means for your role

Developer

Learn one provider SDK well first, then one framework in the layer your work needs most. Keep model calls behind your own small interface so tools can be swapped.

Tester / QA

Ask which layers a feature uses: each layer has its own failure modes, such as retrieval misses, tool errors or prompt regressions, and needs its own tests.

Business Analyst / PO

Use the layer map in requirements discussions: does this feature need retrieval, tools or multi-step workflows? That shapes cost, risk and effort.

Tech Lead / Architect

Standardise one tool per layer for the organisation, document why in an ADR, and review it each year as tools change.

Project / Delivery Manager

Expect framework choices to affect delivery speed and running cost. Budget time for upgrades, since these tools change often.

Scrum Master / Agile

When the team debates tools, steer it to a short time-boxed spike comparing two options on a real task, rather than long opinion debates.

Try it: map an AI feature

  1. Pick an AI feature you use or want to build.
  2. For each layer in the table, write which tool it uses (or 'none needed').
  3. Circle the one layer where a framework would save the most effort, and read that lesson next.

Key takeaways

  • Frameworks are ready-made code for common AI jobs: faster for common patterns, but another thing to learn and depend on.
  • AI applications have layers: models, model access, application frameworks, agents and tools, data and retrieval, quality and operations.
  • Start with the provider SDK; add a framework when it solves a real problem such as retrieval or multi-step agents.
  • Choose by language, fit, maturity, lock-in, licence and your team's platform, and expect tools to change.

Key terms

Framework
Ready-made code that handles common jobs so you build applications faster.
SDK
Software development kit: a provider's official library for calling its service from code.
Stack
The set of tools and services an application is built from.
Open source
Software whose code is public and can be used and changed under its licence.

Check your understanding

Answer all 5 questions. Score 3 or more to complete the lesson.

  1. 1. You need one model call to summarise an email. What should you start with?

  2. 2. Which layer handles searching your documents by meaning?

  3. 3. When does adding a framework usually make sense?

  4. 4. Why prefer tools that work with several model providers?

  5. 5. What is the best source for current code examples of a framework?

Answer all 5 questions to check.

Practise more