From Zero to Building Your Own AI.
A complete, free learning roadmap that takes you from Python and mathematics to neural networks, Transformers, and training your own language models — step by step.
9 Learning Phases
Phase 0 – Phase 8
14 Hands-on Projects
From matrices to LLMs
Free Learning Resources
Open lectures & textbooks
Transformers
Implement scaled dot-product self-attention, multi-head causal masking, RoPE positional encodings, RMSNorm, and SwiGLU blocks.
One Roadmap. From Your First Line of Python to Your First LLM.
Structured into nine progressive phases spanning approximately 78 weeks at the planned study pace. Each phase builds the exact mathematical and engineering prerequisites required by the next.
Phase Overview
Transformers From Scratch
Dissect and implement every component of modern decoder-only Transformer architectures in clean PyTorch.
Core Curriculum Objectives
- Scaled dot-product attention, causal masking, and multi-head projection
- Rotary Positional Embeddings (RoPE), Pre-LN / RMSNorm, and SwiGLU
- Residual stream dynamics, KV caching, and parameter accounting
Project 05: Mini GPT Architecture
Prerequisite: Phase 4 · NLP & Language Modeling
Key Topics Covered
- Python memory model, generators, and object-oriented design
- Vectorized broadcasting and memory layout in NumPy
- Testing, profiling, and reproducible scientific environments
Milestone Deliverable
Vectorized Data & Tensor Utility Toolkit
Complete 9-Phase Architectural Curriculum
Spanning ~78 weeks from initial Python scripts to quantized LLM inference systems.
Learn. Build. Track. Repeat.
Mastery in artificial intelligence comes from pairing mathematical intuition with hands-on implementation. Our four-part methodology eliminates tutorial fatigue.
Step 1 — Follow Your Path
Explore structured AI learning phases arranged in prerequisite order so you never encounter unexplained mathematical symbols or architectural leaps.
Phase 1 · Matrix Calculus & Gradients
Prerequisite for Backpropagation
Phase 3 · Reverse-Mode Autodiff
Prerequisite for Transformer Training
Phase 5 · Causal Self-Attention
Unlocks Pretraining & Alignment
Step 2 — Learn for Free
Access carefully curated free courses, lectures, documentation, and educational resources from leading universities and AI researchers.
MIT 18.06SC Linear Algebra
Full video lectures & problem sets · Free
Karpathy · Neural Networks: Zero to Hero
Code-first deep learning series · Free
Stanford CS336 · Language Modeling
Lecture notes & system assignments · Free
Step 3 — Build Real Projects
Practice what you learn through coding projects, from basic mathematical models to your own Transformer, without hiding behind high-level wrappers.
01 · Pure-Python Matrix Engine
Zero external ML libraries
03 · Scalar & Tensor Autograd Engine
Custom backward() pass verification
05 · Decoder-Only Mini GPT
Causal Multi-Head Attention + RoPE
Step 4 — Track Your Progress
Use interactive checklists and learning milestones to stay organized across long-horizon study sessions.
Derive scaled dot-product attention gradients
Phase 5 · Mathematical Checkpoint
Verify causal mask prevents future token leakage
Phase 5 · Implementation Unit Test
Train 25M parameter model to target validation loss
Phase 6 · Pretraining Milestone
Everything You Need to Go Beyond AI APIs.
Calling an external API takes five lines of code; understanding why gradients stabilize, how tokenizers compress text, and how attention heads route information requires a true engineering workbench.
Structured Roadmap
A progressive curriculum with prerequisites and detailed objectives that connects foundational mathematics directly to modern Transformer engineering.
Curated Free Courses
High-quality free learning resources from institutions and educators such as Harvard, MIT, Stanford, Google, Hugging Face, and Andrej Karpathy.
Hands-on Projects
Practical implementations that teach users how algorithms and architectures work internally—building matrices, autograd, and attention from scratch.
Interactive Checklists
Track completed lessons, modules, courses, and projects with fine-grained topic verification checklists.
Personalized Learning
Plan learning based on study hours and experience level—whether you are starting with Python or skipping ahead to Transformer internals.
Learning Dashboard
Monitor milestones, project progress, and completed learning phases in a calm, distraction-free study workspace.
Don't Just Learn AI. Build It.
Every theoretical phase culminates in a hands-on implementation project. Progress from writing matrix multiplication in pure Python to pretraining your own language model from random weights.
Tier I · Foundational Mathematical & Optimization Engines
Matrix Operations Engine
Implement 2D matrix multiplication, transposes, dot products, and Gaussian elimination in pure Python without external libraries.
Linear Regression From Scratch
Derive and code analytical normal equations alongside batch and mini-batch gradient descent with L2 regularization.
Tier II · Computational Graphs & Deep Learning Internals
Automatic Differentiation Engine
Build a dynamic reverse-mode autodiff engine that tracks operations on a directed acyclic graph and computes exact gradients via the chain rule.
Topological sort · local Jacobian-vector products
GuideNeural Network From Scratch
Construct a multi-layer perceptron with ReLU/GELU activations, He initialization, softmax cross-entropy loss, and momentum optimizers.
Forward caching · vectorized backward pass · weight decay
GuideTier III · Transformer Architecture & LLM Pretraining Capstones
Mini GPT Architecture
Implement a complete decoder-only Transformer with causal multi-head self-attention, rotary embeddings, RMSNorm, and autoregressive sampling.
[B, T, d_model] residual stream · causal mask · KV cache
Your Own Pretrained LLM
Pretrain a custom language model from random initialization on a curated text corpus with gradient accumulation, checkpointing, and perplexity tracking.
End-to-end pretraining loop · validation loss curves
6 Featured Builds Shown · 14 Complete Projects Across All 9 Phases
Explore the complete 14-project engineering platform with interactive checklists, code examples, testing suites, personal notes, and GitHub repository portfolio tracking.
The Best Learning Resources. Without the Paywall.
World-class AI education is already publicly available from leading universities and researchers—if you know what order to study it in. Every recommended resource is 100% free.
Harvard University
CS50's Introduction to Programming with Python
A comprehensive introduction to Python programming, covering functions, variables, control flow, exceptions, libraries, unit testing, file I/O, and object-oriented programming.
MIT OpenCourseWare
Linear Algebra (18.06SC)
A university-level linear algebra course covering vector spaces, matrices, linear systems, eigenvalues, and related topics.
Machine Learning Crash Course
Google's practical, interactive introduction to machine learning fundamentals, linear and logistic regression, gradient descent, generalization, regularization, and neural networks.
Andrej Karpathy
Neural Networks: Zero to Hero
A code-first introduction to neural networks covering automatic differentiation, backpropagation, language models, and GPT implementations.
Hugging Face
Hugging Face LLM Course
Comprehensive open interactive course covering Transformer models, tokenization pipelines, datasets, fine-tuning, and the open-source LLM ecosystem.
Stanford University
CS336 — Language Modeling from Scratch
Graduate-level Stanford course walking through every aspect of building language models from scratch: Transformer architecture, tokenization, pretraining data pipelines, GPU systems, parallelism, and scaling laws.
External course links open directly on the official university or educator portals—no sign-up walls or middleman fees.
Browse Free CoursesYour Learning Journey, Clearly Organized.
Long-horizon technical study succeeds when your next milestone is unmistakable. Open your live learning dashboard to track phases, courses, study sessions, and weekly tasks—or explore the illustrative sample state below.
Phase Completion Progress (Sample State)
Python
Math
Classical ML
Deep Learning
NLP & Tokens
Transformers
Pretraining
Alignment
LLM Systems
Implementing Causal Multi-Head Self-Attention & RoPE
Module 3 of 5 · 68% Complete
Verify causal attention mask tensor shapes on sequence length T=512
Course Progress Indicators
Project Completion Checklist
- Matrix Operations EngineP1
- Linear Regression From ScratchP2
- Automatic Differentiation EngineP3
- Neural Network From ScratchP3
- Mini GPT Architecture (In Progress)P5
- Your Own Pretrained LLMP6
Sample preview state · Track your own 14 milestones in /dashboard.
Weekly Cadence
Target: 10.0 hrs / wk
9.5 hrs (Sample)
Who Is This For?
Whether you are writing your first Python function or already deploying applications with LLM APIs, the curriculum begins with foundational concepts and gradually progresses toward advanced systems.
Complete beginners who know nothing about AI
Background: Starting from zero programming or calculus experience
Starts at Phase 0 with Python fundamentals and builds mathematical intuition step by step before introducing a single neural network layer.
Programmers transitioning into machine learning
Background: Experienced in software engineering, web, or backend systems
Bridges the gap between traditional deterministic software and probabilistic optimization, linear algebra, and tensor computation.
Students interested in AI research
Background: Studying computer science, mathematics, physics, or engineering
Connects textbook derivations directly to clean, readable PyTorch implementations of attention, autograd, and pretraining loops.
Developers who use AI APIs but want to understand the underlying technology
Background: Building applications with prompt engineering or LLM SDKs
Opens the black box—showing how tokenization, KV caching, RoPE, temperature sampling, and fine-tuning actually work under the hood.
Engineers who want to build and train custom language models
Background: Familiar with PyTorch and basic deep learning concepts
Focuses deeply on decoder-only Transformer architecture, data curation, BF16 pretraining, LoRA/DPO alignment, and quantized inference.
This roadmap is a self-paced educational guide focused on genuine mathematical and engineering comprehension—not a commercial bootcamp or credentialing program.
Your First AI Model Starts with Understanding.
No shortcuts. No black boxes. Just a clear path from the fundamentals to building real models.