AI Engineering Learning Roadmap
Complete AI Engineering learning path from LLM fundamentals and prompt engineering through RAG, agents, fine-tuning, and production deployment.
info
Core concepts every AI engineer needs — how LLMs work under the hood.
Overview of AI engineering landscape, key terminology, and what large language models can and cannot do
How autoregressive transformers predict tokens, the pre-training objective, and the scaling laws that drive model performance
How text is split into tokens, subword tokenization algorithms, and the role of embeddings in representing semantic meaning
Transformer architecture deep-dive — attention mechanisms, feed-forward layers, positional encodings, and decoder-only design
Craft effective prompts to get reliable, structured outputs from any model.
Anatomy of a prompt — system messages, user messages, temperature, top-p, and other inference parameters that shape output
Reusable prompt templates and compositional techniques for summarization, extraction, translation, and classification tasks
Prompting models to reason step-by-step through complex problems, reducing hallucination and improving accuracy on math and logic
Providing in-context examples to steer model behavior without fine-tuning, including shot selection and formatting best practices
Techniques for constraining model output to valid JSON, XML, or typed schemas using grammar-based sampling and tool-use patterns
Augment models with your own data using retrieval-augmented generation.
End-to-end retrieval-augmented generation pipeline — indexing, retrieval, contextualization, and generation in a single flow
Semantic vs. fixed-size chunking, overlap windows, and recursive splitting methods for document preprocessing
Indexing and similarity search with vector databases — HNSW, IVF, and hybrid indexes for billion-scale retrieval
Choosing and evaluating embedding models — text-embedding-3-small, voyage, bge, and multilingual embedding trade-offs
Combining dense vector search with sparse keyword matching (BM25) using fusion algorithms like RRF for better relevance
Give models tools, memory, and the ability to take action in the world.
Model Context Protocol — a standardized interface for connecting LLMs to external tools, data sources, and services
Declaring tool schemas that models can invoke, parsing structured arguments, and executing side effects safely
Building autonomous agents that plan, reason, call tools, and iterate on feedback to accomplish complex goals
Coordinating multi-step workflows — routing, state management, retry policies, and human-in-the-loop handoffs
Architecting systems where specialized agents collaborate or compete — delegation, debate, and shared memory patterns
Integrate models into applications, manage costs, and deploy at scale.
OpenAI API reference — chat completions, assistants, embeddings, image generation, and streaming with the official SDKs
Claude API capabilities — extended thinking, tool use, vision, and safety features for building with Anthropic models
Self-hosting Llama, Mistral, Qwen, and other open-weight models — quantization, inference engines, and hardware requirements
Production deployment strategies — scaling, load balancing, auto-scaling GPU instances, and cold-start mitigation
Reducing inference costs through prompt compression, caching, model routing, batching, and choosing appropriate model tiers
Fine-tune custom models, evaluate rigorously, and operate safely in production.
Supervised fine-tuning of base models on custom datasets — LoRA, QLoRA, and full-parameter tuning with modern frameworks
Comparing training paradigms — supervised fine-tuning, direct preference optimization, and reinforcement learning from human feedback
Sourcing, cleaning, deduplicating, and augmenting high-quality datasets for instruction tuning and preference alignment
Systematic evaluation strategies — benchmark suites, human evaluation, LLM-as-judge, and task-specific metrics for quality assurance
Knowledge distillation techniques — training smaller student models from larger teacher models to reduce cost and latency
Implementing content filters, input/output validation, topic restrictions, and adversarial prompt detection for safe deployment
Tracing LLM calls, logging prompts and responses, tracking token usage, alerting on drift, and debugging production issues
Semantic caching of LLM responses, prefix caching, speculative decoding, and batching to minimize time-to-first-token
Managing API rate limits with queuing, retry backoff, token-bucket algorithms, and multi-key rotation strategies
Working with models that process images, audio, and video alongside text — vision encoders, CLIP, and multimodal generation
Advanced reasoning techniques — tree-of-thought, ReAct, tool-augmented reasoning, and self-consistency for complex problem solving
Leveraging LLMs for code synthesis, repair, and refactoring — repository-level context, execution feedback, and test generation
Securing AI systems against prompt injection, data exfiltration, model theft, and supply chain attacks on LLM pipelines
Responsible AI principles — bias mitigation, transparency, fairness auditing, and alignment with human values and regulations
Extend AI capabilities beyond text — images, audio, video, tool use, and emergent reasoning patterns.
Multimodal LLMs, vision encoders, image captioning, visual QA, OCR, diagram understanding, and vision fine-tuning.
Speech-to-text, text-to-speech, voice cloning, audio understanding, and real-time voice interfaces.
Defining tools, JSON schema for function args, parallel tool calls, multi-step tool orchestration, and structured output modes.
Chain-of-thought prompting, self-consistency, tree-of-thought, ReAct loops, and LLM reasoning benchmarks.
Model cards, system cards, red-teaming, constitutional AI, responsible deployment checklists, and regulatory compliance (EU AI Act, Executive Order).