Roadmaps

Frontend Developer
Backend Developer
Database Administrator
Full Stack Developer
Mobile App Developer
DevOps Engineer
AI Engineer
ML Engineer
MLOps Engineer
Network Engineer
Cyber Security Expert
Software Architect

Courses

Loading…

Ad 300×250

Ad 300×600

Ad 320×50

AI Engineer Roadmap

A hands-on roadmap for building AI-powered applications — covers LLM APIs, prompt and context engineering, RAG systems, agents, MCP servers, evaluation, and production deployment.

Ad 728×90

Prerequisites

This roadmap assumes you can already code. Implementing RAG pipelines, AI agents, and MCP servers requires programming ability — Python is the primary language used across the AI ecosystem.

PythonUsed for calling model APIs, building pipelines, agents, and MCP servers.Open roadmap →Back-end DeveloperREST APIs, HTTP, and deployment concepts are assumed in the Working with APIs and Deployment sections.Open roadmap →

0 / 96 topics completed

0%

Junior
01
How LLMs Work?
0/2
- Core LLM Elements
  - Tokens
  - Context
02
AI Models
0/7
- Types of Models
  - closed vs. open
  - self-hosted vs. hosted
- Choosing the Right Model
  - cost
  - latency
  - context
  - privacy
  - capability
03
Platforms & Ecosystem
0/2
- Hugging Face
- Ollama
04
Working with Model APIs
0/4
- Function Calling / Tool Use
- Structured Outputs (JSON mode)
- Streaming Responses
- Prompt Caching
05
Prompt vs. Context Engineering
0/13
- Prompt Engineering
  - Zero-Shot
  - Few-Shot
  - Chain-of-Thought (CoT)
  - ReAct
  - System Prompting
  - Role & Behavior
  - Input Format
  - Structured Output
  - Constraints
- Context Engineering
  - External Memory
  - RAG & Dynamic Retrieval
  - Context Compaction
  - Context Isolation
Mid-Level
06
Embeddings & Vector Databases
0/9
- What are Embeddings?
  - Semantic Search
- Embedding Models
  - Hosted
    - OpenAI
    - Gemini
    - Cohere
  - Open source
    - Sentence Transformers
    - Hugging Face models
- Vector Databases
  - pgvector
  - Pinecone
  - FAISS
07
RAG
0/8
- Implementing RAG
  - Chunking
  - Embedding
  - Vector Database
  - Retrieval
  - Generation
- Ways of Implementing RAG
  - SDKs Directly vs. a framework
  - LangChain
  - LlamaIndex
08
AI Agents
0/10
- Concepts
  - The Agent Loop
    - reason
    - act
    - observe
    - memory & planning
  - Tools & Function Calling
  - ReAct Prompting
  - Multi-Agents
- Building AI Agents
  - Manual Implementation
  - Frameworks/SDKs
    - OpenAI Agents SDK
    - Claude Agent SDK
09
MCP (Model Context Protocol)
0/8
- Core Components
  - Host
  - Client
  - Server
  - Data Layer
  - Transport Layer
- Developing with MCP
  - Building an Server
  - Building an Client
  - Local vs. Remote
10
Multimodal AI
0/8
- Tasks by Modality
  - Vision
    - Image Understanding
    - Image Generation
  - Audio Processing + TTS + STT
- Implementing
  - vision API
  - image-gen API
  - speech API
  - Hugging Face models
    - LangChain
    - LlamaIndex
Senior
11
Evaluation & Observability
0/12
- LLM Evaluations
  - Types
    - Deterministic
    - Model-Based/LLM-as-judge
    - Human
  - Metrics
  - Tools
    - DeepEval
    - RAGAS
- LLM Observability
  - Tracing & Logging
  - Cost/Latency Monitoring
  - Production Monitoring
  - Tools
    - LangSmith
    - Langfuse
- Regression Testing
12
AI Safety, Guardrails & Ethics
0/7
- Safety Issues
  - Prompt Injection
  - Security & Privacy
- Guardrails & Best Practices
  - Content Moderation APIs
  - Input/Output Constraints
  - Adversarial Testing
  - End-user IDs
  - KYC / Use-Case Scoping
13
Deployment & LLMOps
0/6
- Rate Limits
- Retries & Fallbacks
- Cost & Latency Optimization
- Prompt Versioning
- AI-Assisted Coding Tools
  - Claude Code
  - Cursor

Project Checklist

A reference to follow when implementing a real network project.

0 / 26 completed

0%

Design0/5

Model selected based on cost, latency, context window, and task requirements
Context strategy defined — how conversation history is trimmed or summarized near the limit
Structured vs. free-text output decided upfront — JSON mode if downstream parsing is needed
Streaming enabled where response length is unpredictable
Prompt caching evaluated for long static system prompts to reduce cost and latency

Prompt Engineering0/6

System prompt drafted with clear persona, scope, and behavioral constraints
System prompt tested with representative inputs before writing any application code
Few-shot examples added where zero-shot output is inconsistent
Output format specified explicitly — model told what to return and what to avoid
Edge cases covered in the prompt: empty input, off-topic requests, adversarial queries
Prompt versioned in code or a registry — changes are trackable and reversible

Implementation0/5

API client initialized with timeout and retry configuration
Chat history maintained per session and isolated per user — not shared across sessions
Streaming responses piped to the client without buffering the full response first
Structured outputs validated against schema before being consumed downstream
Token usage logged per request for cost attribution

Safety & Guardrails0/5

Input length capped to prevent prompt injection via oversized payloads
System prompt tested against common jailbreak patterns
Output checked for prohibited content before being returned to the user
User IDs passed to the provider where supported for abuse tracking
Rate limiting applied per user or session, not just at the API level

Production0/5

API keys stored in environment variables — never hardcoded or logged
Retry logic with exponential backoff on 429 and 5xx responses
Fallback model configured for when the primary model is unavailable
Every request traced with latency, token count, and model version
Cost per request monitored and alerted on when it spikes

Continue learning with related tracks

Prompt Engineering