Skip to content
Learn AI
বাংলা

Roadmap

AI engineering roadmap

A study path through the open AI Engineering from Scratch curriculum: 20 phases, from setup and math to agents, production and capstone projects. Each lesson opens on the curriculum site. Tick the lessons you finish; your progress stays in this browser.

Learning paths

Not sure where to start? Pick one path. Each path is a short list of lessons from the phases below, in a good order.

  • Start here

    New to AI engineering

    Set up your tools, run the course on your machine, and learn how one lesson works. Then pick a path.

    Start pathGitHub ↗
  • Core path

    Build and ship AI apps

    Prompts, structured output, embeddings and RAG first. Then evals, serving, monitoring and a safe release.

    Start pathGitHub ↗
  • Core path

    Software engineering basics

    The base every AI system needs: repo and environment, debugging, testing, security and release.

    Start pathGitHub ↗
  • Core path

    Work with coding agents

    Give an agent a clear task, plan from the real code, set up its loop, and check its work before you trust it.

    Start pathGitHub ↗
  • Core path

    Product judgment

    Decide what to build: start from the outcome, cut it into small testable slices, measure, and release in steps.

    Start pathGitHub ↗
  • Focused path

    Model Context Protocol (MCP)

    Build MCP servers, make them secure, test them, and run them in production.

    Start pathGitHub ↗
  • Focused path

    Agent skills

    Write skills that agents can load, then test, secure and package them for real agent tools.

    Start pathGitHub ↗
  • Focused path

    Agents, LLMs and MCP under the hood

    How a model turns tokens into a reply, how an agent loop plans and calls tools, and how MCP wires it all to the outside world.

    Start pathGitHub ↗

Open-source repos to learn from →

Lessons done
0
Total lessons
542
Left to study
1187.6 h
Continue reading →

01

Foundations

Tooling, math, classical ML and deep learning. Everything later stands on this.

  1. Phase 0Suggested

    Setup & Tooling

    Set up Python, Git, Docker, GPUs and API keys so every later lesson runs.

    13 lessons · ~15.5 h

    0/13
    13 lessons
    1. 01Dev Environment75 min
    2. 02Git & Collaboration45 min
    3. 03GPU Setup & Cloud75 min
    4. 04APIs & Keys75 min
    5. 05Jupyter Notebooks75 min
    6. 06Python Environments75 min
    7. 07Docker for AI75 min
    8. 08Editor Setup75 min
    9. 09Data Management75 min
    10. 10Terminal & Shell45 min
    11. 11Linux for AI45 min
    12. 12Debugging & Profiling75 min
    13. 13Python for AI Engineering2 h
  2. Phase 1Suggested

    Math Foundations

    Linear algebra, calculus, probability and optimization, the math behind training.

    22 lessons · ~22.5 h

    0/22
    22 lessons
    1. 01Linear Algebra Intuition45 min
    2. 02Vectors, Matrices & Operations75 min
    3. 03Matrix Transformations & Eigenvalues75 min
    4. 04Calculus for ML — Derivatives & Gradients45 min
    5. 05Chain Rule & Automatic Differentiation75 min
    6. 06Probability & Distributions45 min
    7. 07Bayes' Theorem & Statistical Thinking75 min
    8. 08Optimization — Gradient Descent Family75 min
    9. 09Information Theory — Entropy, KL Divergence45 min
    10. 10Dimensionality Reduction — PCA, t-SNE, UMAP75 min
    11. 11Singular Value Decomposition75 min
    12. 12Tensor Operations75 min
    13. 13Numerical Stability45 min
    14. 14Norms & Distances45 min
    15. 15Statistics for ML45 min
    16. 16Sampling Methods75 min
    17. 17Linear Systems75 min
    18. 18Convex Optimization75 min
    19. 19Complex Numbers for AI45 min
    20. 20The Fourier Transform75 min
    21. 21Graph Theory for ML45 min
    22. 22Stochastic Processes45 min
  3. Phase 2Suggested

    ML Fundamentals

    Classical machine learning: regression, trees, evaluation and feature work.

    21 lessons · ~26.3 h

    0/21
    21 lessons
    1. 01What Is Machine Learning — Types & Taxonomy45 min
    2. 02Linear Regression from Scratch75 min
    3. 03Logistic Regression & Classification75 min
    4. 04Decision Trees & Random Forests75 min
    5. 05Support Vector Machines75 min
    6. 06K-Nearest Neighbors & Distance Metrics75 min
    7. 07Unsupervised Learning — K-Means, DBSCAN75 min
    8. 08Feature Engineering & Selection75 min
    9. 09Model Evaluation — Metrics, Cross-Validation75 min
    10. 10Bias, Variance & the Learning Curve45 min
    11. 11Ensemble Methods — Boosting, Bagging, Stacking75 min
    12. 12Hyperparameter Tuning & AutoML75 min
    13. 13ML Pipelines & Experiment Tracking75 min
    14. 14Naive Bayes — Multinomial, Gaussian, Bernoulli75 min
    15. 15Time Series Fundamentals45 min
    16. 16Anomaly Detection75 min
    17. 17Handling Imbalanced Data75 min
    18. 18Feature Selection75 min
    19. 19ML System Design — From Problem to Deployed Model110 min
    20. 20Monitoring Models in Production — Data Drift and Concept Drift110 min
    21. 21Feature Stores and Training-Serving Skew100 min
  4. Phase 3Suggested

    Deep Learning Core

    Neural networks from scratch: backpropagation, layers, regularization and training loops.

    13 lessons · ~14.5 h

    0/13
    13 lessons
    1. 01The Perceptron — Where It All Started45 min
    2. 02Multi-Layer Networks & Forward Pass75 min
    3. 03Backpropagation from Scratch75 min
    4. 04Activation Functions — ReLU, Sigmoid, GELU & Why45 min
    5. 05Loss Functions — MSE, Cross-Entropy, Contrastive45 min
    6. 06Optimizers — SGD, Momentum, Adam, AdamW75 min
    7. 07Regularization — Dropout, Weight Decay, BatchNorm75 min
    8. 08Weight Initialization & Training Stability45 min
    9. 09Learning Rate Schedules & Warmup45 min
    10. 10Build Your Own Mini Framework2 h
    11. 11Introduction to PyTorch75 min
    12. 12Introduction to JAX75 min
    13. 13Debugging Neural Networks75 min

02

Vision, language and speech

How models see, read and hear, before transformers took over.

  1. Phase 4Optional depth

    Computer Vision

    Convolutional networks, detection, segmentation and vision transformers.

    28 lessons · ~31.3 h

    Covered in the Parts: Multimodal Engineering

    0/28
    28 lessons
    1. 01Image Fundamentals — Pixels, Channels, Color Spaces45 min
    2. 02Convolutions from Scratch75 min
    3. 03CNNs — LeNet to ResNet75 min
    4. 04Image Classification75 min
    5. 05Transfer Learning & Fine-Tuning75 min
    6. 06Object Detection — YOLO from Scratch75 min
    7. 07Semantic Segmentation — U-Net75 min
    8. 08Instance Segmentation — Mask R-CNN75 min
    9. 09Image Generation — GANs75 min
    10. 10Image Generation — Diffusion Models75 min
    11. 11Stable Diffusion — Architecture & Fine-Tuning75 min
    12. 12Video Understanding — Temporal Modeling45 min
    13. 133D Vision — Point Clouds, NeRFs45 min
    14. 14Vision Transformers (ViT)45 min
    15. 15Real-Time Vision — Edge Deployment75 min
    16. 16Build a Complete Vision Pipeline2 h
    17. 17Self-Supervised Vision — SimCLR, DINO, MAE75 min
    18. 18Open-Vocabulary Vision — CLIP45 min
    19. 19OCR & Document Understanding45 min
    20. 20Image Retrieval & Metric Learning45 min
    21. 21Keypoint Detection & Pose Estimation45 min
    22. 223D Gaussian Splatting from Scratch90 min
    23. 23Diffusion Transformers & Rectified Flow75 min
    24. 24SAM 3 & Open-Vocabulary Segmentation60 min
    25. 25Vision-Language Models (ViT-MLP-LLM)75 min
    26. 26Monocular Depth & Geometry Estimation60 min
    27. 27Multi-Object Tracking & Video Memory60 min
    28. 28World Models & Video Diffusion75 min
  2. Phase 5Suggested

    NLP — Foundations to Advanced

    Tokenization, embeddings, sequence models and the road to modern language models.

    29 lessons · ~30.5 h

    Covered in the Parts: LLM Fundamentals

    0/29
    29 lessons
    1. 01Text Processing — Tokenization, Stemming, Lemmatization45 min
    2. 02Bag of Words, TF-IDF & Text Representation75 min
    3. 03Word Embeddings — Word2Vec from Scratch75 min
    4. 04GloVe, FastText & Subword Embeddings45 min
    5. 05Sentiment Analysis75 min
    6. 06Named Entity Recognition (NER)75 min
    7. 07POS Tagging & Syntactic Parsing45 min
    8. 08Text Classification — CNNs & RNNs for Text75 min
    9. 09Sequence-to-Sequence Models75 min
    10. 10Attention Mechanism — The Breakthrough45 min
    11. 11Machine Translation75 min
    12. 12Text Summarization75 min
    13. 13Question Answering Systems75 min
    14. 14Information Retrieval & Search75 min
    15. 15Topic Modeling — LDA, BERTopic45 min
    16. 16Text Generation — Language Models Before Transformers45 min
    17. 17Chatbots — Rule-Based to Neural75 min
    18. 18Multilingual NLP45 min
    19. 19Subword Tokenization — BPE, WordPiece, Unigram, SentencePiece60 min
    20. 20Structured Outputs & Constrained Decoding60 min
    21. 21NLI & Textual Entailment60 min
    22. 22Embedding Models Deep Dive60 min
    23. 23Chunking Strategies for RAG60 min
    24. 24Coreference Resolution60 min
    25. 25Entity Linking & Disambiguation60 min
    26. 26Relation Extraction & Knowledge Graph Construction60 min
    27. 27LLM Evaluation — RAGAS, DeepEval, G-Eval75 min
    28. 28Long-Context Evaluation — NIAH, RULER, LongBench, MRCR60 min
    29. 29Dialogue State Tracking75 min
  3. Phase 6Optional depth

    Speech & Audio

    Speech recognition, synthesis and audio models.

    17 lessons · ~18.5 h

    Covered in the Parts: Multimodal Engineering

    0/17
    17 lessons
    1. 01Audio Fundamentals — Waveforms, Sampling, Fourier Transform45 min
    2. 02Spectrograms, Mel Scale & Audio Features45 min
    3. 03Audio Classification75 min
    4. 04Speech Recognition (ASR)45 min
    5. 05Whisper — Architecture & Fine-Tuning75 min
    6. 06Speaker Recognition & Verification45 min
    7. 07Text-to-Speech (TTS)75 min
    8. 08Voice Cloning & Voice Conversion75 min
    9. 09Music Generation75 min
    10. 10Audio-Language Models45 min
    11. 11Real-Time Audio Processing75 min
    12. 12Build a Voice Assistant Pipeline2 h
    13. 13Neural Audio Codecs — EnCodec, SNAC, Mimi, DAC60 min
    14. 14Voice Activity Detection & Turn-Taking45 min
    15. 15Streaming Speech-to-Speech — Moshi, Hibiki75 min
    16. 16Voice Anti-Spoofing & Audio Watermarking75 min
    17. 17Audio Evaluation — WER, MOS, MMAU, Leaderboards60 min

03

Transformers and LLMs

Attention, generative models, reinforcement learning, and a language model built by hand.

  1. Phase 7Suggested

    Transformers Deep Dive

    Attention, the transformer block, positional encoding and the variants in use today.

    16 lessons · ~16.3 h

    Covered in the Parts: LLM Fundamentals

    0/16
    16 lessons
    1. 01Why Transformers — The Problems with RNNs45 min
    2. 02Self-Attention from Scratch75 min
    3. 03Multi-Head Attention75 min
    4. 04Positional Encoding — Sinusoidal, RoPE, ALiBi45 min
    5. 05The Full Transformer — Encoder + Decoder75 min
    6. 06BERT — Masked Language Modeling45 min
    7. 07GPT — Causal Language Modeling75 min
    8. 08T5, BART — Encoder-Decoder Models45 min
    9. 09Vision Transformers (ViT)45 min
    10. 10Audio Transformers — Whisper Architecture45 min
    11. 11Mixture of Experts (MoE)45 min
    12. 12KV Cache, Flash Attention & Inference Optimization75 min
    13. 13Scaling Laws45 min
    14. 14Build a Transformer from Scratch — The Capstone2 h
    15. 15Attention Variants — Sliding Window, Sparse, Differential60 min
    16. 16Speculative Decoding — Draft, Verify, Repeat60 min
  2. Phase 8Optional depth

    Generative AI

    Generative models: autoencoders, GANs, diffusion and flow-based methods.

    15 lessons · ~15.5 h

    Covered in the Parts: Multimodal Engineering

    0/15
    15 lessons
    1. 01Generative Models — Taxonomy & History45 min
    2. 02Autoencoders & VAE75 min
    3. 03GANs — Generator vs Discriminator75 min
    4. 04Conditional GANs & Pix2Pix75 min
    5. 05StyleGAN45 min
    6. 06Diffusion Models — DDPM from Scratch75 min
    7. 07Latent Diffusion & Stable Diffusion75 min
    8. 08ControlNet, LoRA & Image Conditioning75 min
    9. 09Inpainting, Outpainting & Image Editing75 min
    10. 10Video Generation45 min
    11. 11Audio Generation45 min
    12. 123D Generation45 min
    13. 13Flow Matching & Rectified Flows45 min
    14. 14Evaluation — FID, CLIP Score, Human Preference45 min
    15. 19Visual Autoregressive Modeling (VAR): Next-Scale Prediction90 min
  3. Phase 9Optional depth

    Reinforcement Learning

    Reinforcement learning, from bandits and Q-learning to policy gradients.

    12 lessons · ~13 h

    0/12
    12 lessons
    1. 01MDPs, States, Actions & Rewards45 min
    2. 02Dynamic Programming75 min
    3. 03Monte Carlo Methods75 min
    4. 04Temporal Difference — Q-Learning, SARSA75 min
    5. 05Deep Q-Networks (DQN)75 min
    6. 06Policy Gradient Methods — REINFORCE75 min
    7. 07Actor-Critic — A2C, A3C75 min
    8. 08Proximal Policy Optimization (PPO)75 min
    9. 09Reward Modeling & RLHF45 min
    10. 10Multi-Agent RL45 min
    11. 11Sim-to-Real Transfer45 min
    12. 12RL for Games75 min
  4. Phase 10Suggested

    LLMs from Scratch

    Pre-train, align and sample from a language model you wrote yourself.

    24 lessons · ~28.4 h

    Covered in the Parts: LLM Fundamentals, Inference

    0/24
    24 lessons
    1. 01Tokenizers — BPE, WordPiece, SentencePiece45 min
    2. 02Building a Tokenizer from Scratch75 min
    3. 03Data Pipelines for Pre-Training75 min
    4. 04Pre-Training a Mini GPT (124M)2 h
    5. 05Scaling — Distributed Training, FSDP, DeepSpeed75 min
    6. 06Instruction Tuning — SFT75 min
    7. 07RLHF — Reward Model + PPO Training75 min
    8. 08DPO — Direct Preference Optimization75 min
    9. 09Constitutional AI & Self-Improvement45 min
    10. 10Evaluation — Benchmarks, Evals, LM Harness75 min
    11. 11Quantization — INT8, GPTQ, AWQ, GGUF75 min
    12. 12Inference Optimization75 min
    13. 13Building a Complete LLM Pipeline2 h
    14. 14Open Models — Architecture Walkthroughs45 min
    15. 15Speculative Decoding and EAGLE-375 min
    16. 16Differential Attention (V2)60 min
    17. 17Native Sparse Attention (DeepSeek NSA)60 min
    18. 18Multi-Token Prediction (MTP)60 min
    19. 19DualPipe Parallelism60 min
    20. 20DeepSeek-V3 Architecture Walkthrough75 min
    21. 21Jamba — Hybrid SSM-Transformer60 min
    22. 22Async and Hogwild! Inference60 min
    23. 25Speculative Decoding and EAGLE75 min
    24. 34Gradient Checkpointing and Activation Recomputation70 min

04

Building with LLMs

Prompts, retrieval, structured output, multimodal inputs, tools and protocols such as MCP.

  1. Phase 11Suggested

    LLM Engineering

    The day-to-day craft: prompts, structured outputs, RAG, tool use, evaluation, caching and guardrails.

    17 lessons · ~18.8 h

    Covered in the Parts: Prompt Engineering, Context Engineering, RAG and Knowledge Systems, Tool Use and Integrations, Evaluation Engineering, Cost and Performance Optimization, Safety, Security, and Guardrails

    0/17
    17 lessons
    1. 01Prompt Engineering — Techniques & Patterns45 min
    2. 02Few-Shot, Chain-of-Thought, Tree-of-Thought45 min
    3. 03Structured Outputs75 min
    4. 04Embeddings & Vector Representations75 min
    5. 05Context Engineering75 min
    6. 06RAG — Retrieval-Augmented Generation75 min
    7. 07Advanced RAG75 min
    8. 08Fine-Tuning with LoRA & QLoRA75 min
    9. 09Function Calling & Tool Use75 min
    10. 10Evaluation & Testing LLM Applications45 min
    11. 11Caching, Rate Limiting & Cost Optimization45 min
    12. 12Guardrails, Safety & Content Filtering45 min
    13. 13Building a Production LLM Application2 h
    14. 14Model Context Protocol (MCP)75 min
    15. 15Prompt Caching & Context Caching60 min
    16. 16Agent State Machines — Graphs, Nodes, Checkpoints75 min
    17. 17Agent Framework Tradeoffs45 min
  2. Phase 12Optional depth

    Multimodal AI

    Models that take images, audio and video alongside text.

    25 lessons · ~67 h

    Covered in the Parts: Multimodal Engineering

    0/25
    25 lessons
    1. 01Vision Transformers and the Patch-Token Primitive2 h
    2. 02CLIP and Contrastive Vision-Language Pretraining3 h
    3. 03BLIP-2 and Q-Former as Modality Bridge3 h
    4. 04Flamingo and Gated Cross-Attention2 h
    5. 05LLaVA and Visual Instruction Tuning3 h
    6. 06Any-Resolution Vision: Patch-n'-Pack and NaFlex2 h
    7. 07Open-Weight VLM Recipes: What Actually Matters3 h
    8. 08LLaVA-OneVision: Single, Multi, Video3 h
    9. 09Qwen-VL Family and Dynamic-FPS Video2 h
    10. 10InternVL3 Native Multimodal Pretraining2 h
    11. 11Chameleon and Early-Fusion Token-Only3 h
    12. 12Emu3 Next-Token Prediction for Generation2 h
    13. 13Transfusion Autoregressive + Diffusion3 h
    14. 14Show-o and Discrete-Diffusion Unified2 h
    15. 15Janus-Pro Decoupled Encoders2 h
    16. 16MIO Any-to-Any Streaming2 h
    17. 17Video-Language Temporal Grounding3 h
    18. 18Long-Video Understanding at Million-Token Context3 h
    19. 19Audio-Language Models: Whisper to AF33 h
    20. 20Omni Models: Thinker-Talker3 h
    21. 21Embodied VLAs: RT-2, OpenVLA, π0, GR00T3 h
    22. 22Document and Diagram Understanding3 h
    23. 23ColPali Vision-Native Document RAG3 h
    24. 24Multimodal RAG and Cross-Modal Retrieval3 h
    25. 25Multimodal Agents and Computer-Use (Capstone)4 h
  3. Phase 13Suggested

    Tools & Protocols

    Function calling, MCP and the protocols that connect models to tools and to each other.

    45 lessons · ~75.7 h

    Covered in the Parts: Tool Use and Integrations

    0/45
    45 lessons
    1. 01The Tool Interface45 min
    2. 02Function Calling Deep Dive75 min
    3. 03Parallel and Streaming Tool Calls75 min
    4. 04Structured Output75 min
    5. 05Tool Schema Design45 min
    6. 06MCP Fundamentals: Stateless Requests and JSON-RPC55 min
    7. 07Building an MCP Server: Stateless Python and TypeScript85 min
    8. 08Building an MCP Client: Discovery, Routing, and Dual-Era Fallback85 min
    9. 09MCP Transports: stdio and Stateless Streamable HTTP65 min
    10. 10MCP Resources and Prompts: Addressable Context for Stateless Servers60 min
    11. 11MCP Model Input: Sampling Migration and Stateless MRTR75 min
    12. 12Explicit Scope and Stateless Elicitation60 min
    13. 13MCP Tasks Extension: Durable Work on a Stateless Core90 min
    14. 14MCP Apps on the Stateless Protocol75 min
    15. 15MCP Security: Poisoned Metadata, Routing, and MRTR State60 min
    16. 16MCP Authorization: CIMD, Issuer Binding, PKCE, and Step-Up90 min
    17. 17Stateless MCP Gateways and Registry Admission75 min
    18. 18MCP Auth in Production: Issuer-Bound Enrollment and Tokens90 min
    19. 19A2A Protocol75 min
    20. 20OpenTelemetry GenAI75 min
    21. 21LLM Routing Layer45 min
    22. 22Agent Skills: Portable Contract and Runtime Boundary90 min
    23. 23Capstone: Stateless Tool Ecosystem2 h
    24. 24Skill Discovery and Progressive Disclosure105 min
    25. 25Skill Invocation and Routing105 min
    26. 26Skill Permissions, Sandboxes, and Trust2 h
    27. 27Skill Evals, Packaging, and Portability2.5 h
    28. 28MCP Tool Contracts and Content2 h
    29. 29MCP Reliability, Cancellation, and Flow Control2 h
    30. 30MCP Registry Supply Chain: Admission, Drift, and Rollback90 min
    31. 31MCP Conformance Engineering: Versioning, Evidence, and Operations100 min
    32. 32Program Anatomy: Reading ELF From Scratch2 h
    33. 33Disassembly and Control-Flow Graphs2.5 h
    34. 34Lifting Machine Code to Pseudocode2.5 h
    35. 35Evidence, Confidence, and Unknowns in Binary Analysis105 min
    36. 36Rebuild and Prove Equal110 min
    37. 37The Reverse Engineering Landscape2.5 h
    38. 38Calling Conventions and Stack Frames2.3 h
    39. 39Compiler Idioms and Optimization2.5 h
    40. 40Linking, Loading, and Stripped Binaries2.7 h
    41. 41Recovering Data and Types2.3 h
    42. 42Dynamic Analysis of Your Own Program2.2 h
    43. 43Bytecode and Managed Formats2.5 h
    44. 44Binary Diffing and Symbolic Execution2.5 h
    45. 45AI-Assisted Reverse Engineering2.3 h

05

Agents

Agent loops, long-running autonomous systems, and teams of agents.

  1. Phase 14Suggested

    Agent Engineering

    Build agent runtimes: loops, planning, memory, state machines and evaluation.

    54 lessons · ~56.4 h

    Covered in the Parts: Agents and Agentic Systems, Context Engineering

    0/54
    54 lessons
    1. 01The Agent Loop60 min
    2. 02ReWOO and Plan-and-Execute60 min
    3. 03Reflexion and Verbal Reinforcement Learning60 min
    4. 04Tree of Thoughts and LATS75 min
    5. 05Self-Refine and CRITIC60 min
    6. 06Tool Use and Function Calling60 min
    7. 07Agent Memory — Virtual Context and Memory Paging75 min
    8. 08Memory Blocks and Sleep-Time Compute75 min
    9. 09Hybrid Memory — Vector + Graph + KV75 min
    10. 10Skill Libraries and Lifelong Learning (Voyager)75 min
    11. 11Planning with HTN and Evolutionary Search75 min
    12. 12Anthropic's Workflow Patterns60 min
    13. 13Stateful Graph Orchestration — Durable Execution and Checkpoints75 min
    14. 14The Actor Model for Agents75 min
    15. 15Role-Based Agent Teams — Roles, Tasks, Processes60 min
    16. 16OpenAI Agents SDK — Handoffs, Guardrails, Tracing75 min
    17. 17The Harness as a Library — Subagents and Session Store75 min
    18. 18Production Agent Runtimes45 min
    19. 19Benchmarks — SWE-bench, GAIA, AgentBench60 min
    20. 20Benchmarks — WebArena and OSWorld60 min
    21. 21Computer Use — Claude, OpenAI CUA, Gemini60 min
    22. 22Voice Agents — Pipecat and LiveKit60 min
    23. 23OpenTelemetry GenAI Semantic Conventions60 min
    24. 24Agent Observability — Langfuse, Phoenix, Opik45 min
    25. 25Multi-Agent Debate and Collaboration60 min
    26. 26Failure Modes — Why Agents Break60 min
    27. 27Prompt Injection and the PVE Defense75 min
    28. 28Orchestration Patterns — Supervisor, Swarm, Hierarchical60 min
    29. 29Production Runtimes — Queue, Event, Cron60 min
    30. 30Eval-Driven Agent Development60 min
    31. 31Agent Workbench: Why Capable Models Still Fail45 min
    32. 32The Minimal Agent Workbench45 min
    33. 33Agent Instructions as Executable Constraints50 min
    34. 34Repo Memory and Durable State60 min
    35. 35Initialization Scripts for Agents45 min
    36. 36Scope Contracts and Task Boundaries50 min
    37. 37Runtime Feedback Loops50 min
    38. 38Verification Gates55 min
    39. 39Reviewer Agent: Separate Builder from Marker55 min
    40. 40Multi-Session Handoff50 min
    41. 41The Workbench on a Real Repo60 min
    42. 42Capstone: Ship a Reusable Agent Workbench Pack75 min
    43. 43Frame the Task Before the Agent Writes Code60 min
    44. 44Build an Evidence-Backed Execution Plan65 min
    45. 45Delegate Agent Work with Isolation and Merge Contracts70 min
    46. 46Turn Every Agent Correction into a System Improvement65 min
    47. 47Define the Outcome Before You Choose the Output60 min
    48. 48Discover the Workflow People Actually Perform70 min
    49. 49Map Assumptions and Resolve the Riskiest One First65 min
    50. 50Choose the Smallest Slice That Can Change the Decision65 min
    51. 51Write Specifications That Preserve Judgment75 min
    52. 52Design Success Metrics Before the Result Exists70 min
    53. 53Choose Prototype, Pilot, or Production Deliberately70 min
    54. 54Build a Feedback Ratchet with Ownership and Retirement75 min
  2. Phase 15Suggested

    Autonomous Systems

    Long-running agents that act on their own, and the controls that keep them in bounds.

    22 lessons · ~20 h

    Covered in the Parts: Agents and Agentic Systems

    0/22
    22 lessons
    1. 01From Chatbots to Long-Horizon Agents (METR)45 min
    2. 02STaR, V-STaR, Quiet-STaR — Self-Taught Reasoning60 min
    3. 03AlphaEvolve — Evolutionary Coding Agents60 min
    4. 04Darwin Gödel Machine — Self-Modifying Agents60 min
    5. 05AI Scientist v2 — Workshop-Level Research60 min
    6. 06Automated Alignment Research (Anthropic AAR)60 min
    7. 07Recursive Self-Improvement — Capability vs Alignment60 min
    8. 08Bounded Self-Improvement Designs60 min
    9. 09Autonomous Coding Agent Landscape (SWE-bench, CodeAct)45 min
    10. 10Permission Modes for Autonomous Agents45 min
    11. 11Browser Agents and Indirect Prompt Injection45 min
    12. 12Durable Execution for Long-Running Agents60 min
    13. 13Action Budgets, Iteration Caps, Cost Governors60 min
    14. 14Kill Switches, Circuit Breakers, Canary Tokens60 min
    15. 15HITL — Propose-Then-Commit60 min
    16. 16Checkpoints and Rollback60 min
    17. 17Constitutional AI and Rule Overrides60 min
    18. 18Llama Guard and Input/Output Classification45 min
    19. 19Anthropic Responsible Scaling Policy v3.045 min
    20. 20OpenAI Preparedness Framework and DeepMind FSF45 min
    21. 21METR Time Horizons and External Evaluation60 min
    22. 22CAIS, CAISI, and Societal-Scale Risk45 min
  3. Phase 16Suggested

    Multi-Agent & Swarms

    Several agents working together: coordination, handoffs and swarm patterns.

    25 lessons · ~29.3 h

    Covered in the Parts: Agents and Agentic Systems

    0/25
    25 lessons
    1. 01Why Multi-Agent45 min
    2. 02FIPA-ACL Heritage and Speech Acts60 min
    3. 03Communication Protocols45 min
    4. 04The Multi-Agent Primitive Model60 min
    5. 05Supervisor / Orchestrator-Worker Pattern75 min
    6. 06Hierarchical Architecture and Decomposition Drift60 min
    7. 07Society of Mind and Multi-Agent Debate75 min
    8. 08Role Specialization — Planner / Critic / Executor / Verifier75 min
    9. 09Parallel Swarm and Networked Architectures60 min
    10. 10Group Chat and Speaker Selection60 min
    11. 11Handoffs and Routines (Stateless Orchestration)60 min
    12. 12A2A — The Agent-to-Agent Protocol75 min
    13. 13Shared Memory and Blackboard Patterns75 min
    14. 14Consensus and Byzantine Fault Tolerance for Agents75 min
    15. 15Voting, Self-Consistency, and Debate Topology75 min
    16. 16Negotiation and Bargaining75 min
    17. 17Generative Agents and Emergent Simulation75 min
    18. 18Theory of Mind and Emergent Coordination75 min
    19. 19Swarm Optimization for LLMs (PSO, ACO)75 min
    20. 20MARL — MADDPG, QMIX, MAPPO90 min
    21. 21Agent Economies, Token Incentives, Reputation75 min
    22. 22Production Scaling — Queues, Checkpoints, Durability75 min
    23. 23Failure Modes — MAST, Groupthink, Monoculture, Cascading75 min
    24. 24Evaluation and Coordination Benchmarks75 min
    25. 25Case Studies and 2026 State of the Art90 min

06

Production and safety

Serving, observability and cost, then ethics, safety and alignment.

  1. Phase 17Suggested

    Infrastructure & Production

    Serving, scaling, observability and cost control for AI systems in production.

    29 lessons · ~30.5 h

    Covered in the Parts: Inference, LLMOps and Observability, Cost and Performance Optimization, AI Application Architecture

    0/29
    29 lessons
    1. 01Managed LLM Platforms — Bedrock, Azure OpenAI, Vertex AI60 min
    2. 02Inference Platform Economics — Fireworks, Together, Baseten, Modal60 min
    3. 03GPU Autoscaling on Kubernetes — Karpenter, KAI Scheduler75 min
    4. 04Serving Engine Internals — PagedAttention, Continuous Batching, Chunked Prefill75 min
    5. 05EAGLE-3 Speculative Decoding in Production60 min
    6. 06Prefix-Cache Serving — RadixAttention and KV Reuse60 min
    7. 07Hardware-Specialized Inference Compilation — FP8 and NVFP4 on Blackwell75 min
    8. 08Inference Metrics — TTFT, TPOT, ITL, Goodput, P9960 min
    9. 09Production Quantization — AWQ, GPTQ, GGUF, FP8, NVFP475 min
    10. 10Cold Start Mitigation for Serverless LLMs60 min
    11. 11Multi-Region LLM Serving and KV Cache Locality60 min
    12. 12Edge Inference — ANE, Hexagon, WebGPU, Jetson60 min
    13. 13LLM Observability Stack Selection60 min
    14. 14Prompt Caching and Semantic Caching Economics60 min
    15. 15Batch APIs — the 50% Discount as Industry Standard45 min
    16. 16Model Routing as a Cost-Reduction Primitive60 min
    17. 17Disaggregated Prefill/Decode — NVIDIA Dynamo and llm-d75 min
    18. 18Production Serving Stack — KV Offloading and Cache-Aware Routing60 min
    19. 19AI Gateways — LiteLLM, Portkey, Kong, Bifrost60 min
    20. 20Shadow, Canary, and Progressive Deployment60 min
    21. 21A/B Testing LLM Features — GrowthBook and Statsig60 min
    22. 22Load Testing LLM APIs — k6, LLMPerf, GenAI-Perf75 min
    23. 23SRE for AI — Multi-Agent Incident Response60 min
    24. 24Chaos Engineering for LLM Production60 min
    25. 25Security — Secrets, PII Scrubbing, Audit Logs60 min
    26. 26Compliance — SOC 2, HIPAA, GDPR, EU AI Act, ISO 4200160 min
    27. 27FinOps for LLMs — Unit Economics and Multi-Tenant Attribution60 min
    28. 28Self-Hosted Serving Selection — Matching Engine to Hardware and Scale45 min
    29. 29Model Demo Interfaces — From a Function to a Shareable App90 min
  2. Phase 18Suggested

    Ethics, Safety & Alignment

    Ethics, safety, red-teaming, security and alignment.

    30 lessons · ~30.8 h

    Covered in the Parts: Safety, Security, and Guardrails, Evaluation Engineering

    0/30
    30 lessons
    1. 01Instruction-Following as Alignment Signal45 min
    2. 02Reward Hacking & Goodhart's Law60 min
    3. 03Direct Preference Optimization Family60 min
    4. 04Sycophancy as RLHF Amplification45 min
    5. 05Constitutional AI & RLAIF60 min
    6. 06Mesa-Optimization & Deceptive Alignment75 min
    7. 07Sleeper Agents — Persistent Deception60 min
    8. 08In-Context Scheming in Frontier Models60 min
    9. 09Alignment Faking60 min
    10. 10AI Control — Safety Despite Subversion75 min
    11. 11Scalable Oversight & Weak-to-Strong Generalization60 min
    12. 12Red-Teaming — PAIR & Automated Attacks75 min
    13. 13Many-Shot Jailbreaking45 min
    14. 14ASCII Art & Visual Jailbreaks60 min
    15. 15Indirect Prompt Injection75 min
    16. 16Red-Team Tooling — Garak, Llama Guard, PyRIT75 min
    17. 17WMDP & Dual-Use Capability Evaluation60 min
    18. 18Frontier Safety Frameworks — RSP, PF, FSF75 min
    19. 19Model Welfare Research45 min
    20. 20Bias & Representational Harm60 min
    21. 21Fairness Criteria — Group, Individual, Counterfactual60 min
    22. 22Differential Privacy for LLMs60 min
    23. 23Watermarking — SynthID, Stable Signature, C2PA75 min
    24. 24Regulatory Frameworks — EU, US, UK, Korea75 min
    25. 25EchoLeak & CVEs for AI45 min
    26. 26Model, System & Dataset Cards60 min
    27. 27Data Provenance & Training-Data Governance60 min
    28. 28Alignment Research Ecosystem — MATS, Redwood, Apollo, METR45 min
    29. 29Moderation Systems — OpenAI, Perspective, Llama Guard60 min
    30. 30Dual-Use Risk — Cyber, Bio, Chem, Nuclear75 min

07

Capstone

Full projects that tie the earlier phases together.

  1. Phase 19Suggested

    Capstone Projects

    Capstone projects, from a terminal coding agent to a stateless MCP server.

    85 lessons · ~627 h

    Covered in the Parts: AI Application Architecture

    0/85
    85 lessons
    1. 01Terminal-Native Coding Agent35 h
    2. 02RAG over Codebase (Cross-Repo Semantic Search)30 h
    3. 03Real-Time Voice Assistant (ASR to LLM to TTS)30 h
    4. 04Multimodal Document QA (Vision-First)30 h
    5. 05Autonomous Research Agent (AI-Scientist Class)40 h
    6. 06DevOps Troubleshooting Agent for Kubernetes30 h
    7. 07End-to-End Fine-Tuning Pipeline35 h
    8. 08Production RAG Chatbot (Regulated Vertical)30 h
    9. 09Code Migration Agent (Repo-Level Upgrade)30 h
    10. 10Multi-Agent Software Engineering Team40 h
    11. 11LLM Observability & Eval Dashboard25 h
    12. 12Video Understanding Pipeline (Scene to QA)30 h
    13. 13Stateless MCP Server with Registry and Governance25 h
    14. 14Speculative-Decoding Inference Server30 h
    15. 15Constitutional Safety Harness + Red-Team Range25 h
    16. 16GitHub Issue-to-PR Autonomous Agent30 h
    17. 17Personal AI Tutor (Adaptive, Multimodal)30 h
    18. 20Agent Harness Loop Contract90 min
    19. 21Tool Registry with Schema Validation90 min
    20. 22JSON-RPC 2.0 Over Newline-Delimited Stdio90 min
    21. 23Function Call Dispatcher90 min
    22. 24Plan-Execute Control Flow90 min
    23. 25Verification Gates and the Observation Budget90 min
    24. 26Sandbox Runner with Denylist and Path Jail90 min
    25. 27Eval Harness with Fixture Tasks90 min
    26. 28Observability with OTel GenAI Spans and Prometheus Metrics90 min
    27. 29End-to-End Coding Agent on the Harness90 min
    28. 30BPE Tokenizer From Scratch90 min
    29. 31Tokenized Dataset with Sliding Window90 min
    30. 32Token and Positional Embeddings90 min
    31. 33Multi-Head Self-Attention90 min
    32. 34Transformer Block from Scratch90 min
    33. 35GPT Model Assembly90 min
    34. 36Training Loop and Evaluation90 min
    35. 37Loading Pretrained Weights90 min
    36. 38Classifier Fine-Tuning by Head Swap90 min
    37. 39Instruction Tuning by Supervised Fine-Tuning90 min
    38. 40Direct Preference Optimization from Scratch90 min
    39. 41Full Evaluation Pipeline90 min
    40. 42Large Corpus Downloader90 min
    41. 43HDF5 Tokenized Corpus90 min
    42. 44Cosine LR with Linear Warmup90 min
    43. 45Gradient Clipping and Mixed Precision90 min
    44. 46Gradient Accumulation90 min
    45. 47Checkpoint Save and Resume90 min
    46. 48Distributed Data Parallel and FSDP from Scratch90 min
    47. 49Language Model Evaluation Harness90 min
    48. 50Hypothesis Generator90 min
    49. 51Literature Retrieval90 min
    50. 52Experiment Runner90 min
    51. 53Result Evaluator90 min
    52. 54Paper Writer90 min
    53. 55Critic Loop90 min
    54. 56Iteration Scheduler90 min
    55. 57End-to-End Research Demo90 min
    56. 58Vision Encoder Patches90 min
    57. 59Vision Transformer Encoder90 min
    58. 60Projection Layer for Modality Alignment90 min
    59. 61Cross-Attention Fusion90 min
    60. 62Vision-Language Pretraining90 min
    61. 63Multimodal Evaluation90 min
    62. 64Chunking Strategies, Compared90 min
    63. 65Hybrid Retrieval with BM25 and Dense Embeddings90 min
    64. 66Cross-Encoder Reranker90 min
    65. 67Query Rewriting: HyDE, Multi-Query, and Decomposition90 min
    66. 68RAG Evaluation: Precision, Recall, MRR, nDCG, Faithfulness, Answer Relevance90 min
    67. 69End-to-End RAG System90 min
    68. 70Task Spec Format90 min
    69. 71Classical Metrics90 min
    70. 72Code Exec Metric90 min
    71. 73Perplexity and Calibration90 min
    72. 74Leaderboard Aggregation90 min
    73. 75End-to-End Eval Runner90 min
    74. 76Collective Ops From Scratch90 min
    75. 77Data Parallel DDP From Scratch90 min
    76. 78ZeRO Optimizer State Sharding90 min
    77. 79Pipeline Parallel and Bubble Analysis90 min
    78. 80Sharded Checkpoint and Atomic Resume90 min
    79. 81End-to-End Distributed Training90 min
    80. 82Jailbreak Taxonomy90 min
    81. 83Prompt Injection Detector90 min
    82. 84Refusal Evaluation90 min
    83. 85Content Classifier Integration90 min
    84. 86Constitutional Rules Engine90 min
    85. 87End-to-End Safety Gate90 min

Source: the ROADMAP.md of the curriculum. Lesson titles stay in English. GitHub ↗