AI Engineering Roadmap

Tokenization, context windows, few-shot prompting, structured outputs.
Tokenization- track token usage and latency
Context window - how much info model can consider, what happen if they exceed it
Temperature and other parameter - how affect randomness and creativity
Structured outputs - clean json, response format, output validation, handle when model ignore instrution
Streaming generation, error handling when api fail, rate limits, retry logic, cost management
Tool use (function calling), tool chaining. Build multi-step applications with real actions.
Tools Schema, Augment Validation, Failure handling, Guardrails
Zero-shot, few-shot prompting
• Chain-of-thought prompting
• KV cache and how it works
• Prompt caching: what it is and when it cuts costs

Course

Transformer in Practice (VP @ AMD 🧠)

https://www.deeplearning.ai/courses/transformers-in-practice

Hugging Face (must)

https://www.youtube.com/watch?v=XNDY7jtihok&t=7910s

After hands-on above, do Full Course here, with advanced projects

https://www.learnhuggingface.com/

LangChain

https://www.deeplearning.ai/courses/langchain

Advance

https://www.deeplearning.ai/courses/functions-tools-agents-langchain

Project-

Domain specific assistance, structured data extractor.

Research assistance that search's web, read files, call an API.

AI Engineering Book by Chip Huyen (Bible for AI Engineering)

Only book you need to read for AI Engineering

Link: https://drive.google.com/file/d/1jfn7vc8yVW7S-fToWVotjeRYuoildK29/view?usp=sharing

Latest Workshop and Event in AI Engineering

https://www.youtube.com/@aiDotEngineer

RAG Systems

Learn about-

Chunking, embeddings, vector databases, retrieval strategies.
What retrieval-augmented generation actually solves
Chunking strategies, avoid retrival modes, configure mode and embedding models.
Context assembly problem
Retrieval strategies - similarity, hybrid, reranking, query transformation

Learn these Framework

Pgvector - index and distance metrics choises, hybrid retrieval pattens, Chunking failure modes
weaviate - Hybrid retrival, schema design, ingestion pattens, index configuration pattens, multi-tetants pattens

Course (weaviate included)

https://www.deeplearning.ai/courses/retrieval-augmented-generation

pgvector
https://www.youtube.com/watch?v=hAdEuDBN57g

Prompt Compression and Query Optimization

https://www.deeplearning.ai/courses/prompt-compression-and-query-optimization

AI Agent

Learn about-

Agent Loops & Protocols
Core agent loop, planning, memory, error handling.
Agent Swarms, Self improving Agents, RL for Agents.
State Machine, Planning vs execution, memory management, obserbility.
Chain of thought, Tree of thought.
Episodic vs semantic memory. Red teaming. Cost Saving.

Tool use and function calling

ReAct pattern: Reasoning + action (Thought → Action → Observation loop)
Multi-step planning and agent loops
Where agents break in production.

Build your own Claude Code Agent with Python: Learn about LLM APIs, tool calling, agent loops and more

https://app.codecrafters.io/courses/claude-code/overview

LangGraph (must)

https://www.deeplearning.ai/courses/long-term-agentic-memory-with-langgraph

Pydantic for LLM Workflows

https://www.deeplearning.ai/courses/pydantic-for-llm-workflows

Building Your Own Database Agent

https://www.deeplearning.ai/courses/building-your-own-database-agent

Building and Evaluating Data Agents

https://www.deeplearning.ai/courses/building-and-evaluating-data-agents

Project-

Multi turn conversation app (conversation history, context management, iteration workflow)

Multi AI Agent (Optional)

https://www.deeplearning.ai/courses/practical-multi-ai-agents-and-advanced-use-cases-with-crewai

Voice Agents (Optional)

https://www.deeplearning.ai/courses/building-ai-voice-agents-for-production

Google’s ADK Live Voice Agents (Optional)

https://www.deeplearning.ai/courses/building-live-voice-agents-with-googles-adk

MCP

Least privilege tool exposure, auditing

Start here

https://anthropic.skilljar.com/introduction-to-model-context-protocol

Advance

https://anthropic.skilljar.com/model-context-protocol-advanced-topics

Claude Code in CLI

Start here

https://anthropic.skilljar.com/claude-code-in-action

Claude Sub-agents

https://anthropic.skilljar.com/introduction-to-subagents

Agent Skills

https://anthropic.skilljar.com/introduction-to-agent-skills

Evals

Learn about-

Metrics like Faithfulness, relevance, groundedness. Find which is relevent to your use case and why.
Build dataset from real dataset. Run offline vs online evals workflows. Use diff evaluator types - rule based/LLM as a judge/human review.
Build evaluation sets, Run regression test on prompt, track latency, cost and error rates. Set up alerts.
Regression testing for retrival and prompt changes.
Observability, Tracing and monitoring. Trace level debugging, comparing releases on latency, cost and quality. Linking evaluation regression back to concrete traces.

Learn these Framework

Langsmith - Native observability and evaluation.
ragas - LLM-as-a-Judge

Deployment and Production

Learn about-

FastAPI, Docker, cloud deployment, logging, observability, cost tracking, caching.
Container, Async execution, background jobs, latency monitoring.
Cost tracking - per request, day, user. Caching aggressively, batch request, use smaller models when possible. Health checks, basic alerting, optimize cost.
High performance Inference. For embedding, rerankers, cv models.
Nvidia Triton server, Bento ML. Dynamic patching, concurancy tuning, model repository, Ensembles for pipeline serving.

Google LLMOps

https://www.deeplearning.ai/courses/llmops