Phase 0: Foundations
2–3 Weeks
Prerequisite
Beginner
Before writing a single line of ML code, set up your environment and understand what AI actually is. This phase prevents the confusion that derails most beginners. Rushing past it causes pain later.
Learning Objectives
- Understand the AI landscape: AI vs ML vs DL vs GenAI vs Agentic AI
- Set up Python environment (Anaconda, pip, virtual environments)
- Get comfortable with Jupyter Notebooks for interactive coding
- Learn Git and push your first project to GitHub
- Create accounts on Kaggle and Hugging Face
Topics
Must Know
- What is AI/ML/DL/GenAI — the landscape
- Python setup — Anaconda, pip, venv
- Jupyter Notebooks — interactive coding environment
- Git + GitHub — version control fundamentals
- Command Line basics — terminal navigation
Good to Know
- VS Code setup — Python extensions, shortcuts
- Kaggle account — competitions and datasets
- Hugging Face account — model hub access
- Google Colab — free GPU access for training
Best Free Resources
3Blue1Brown — Neural Networks Playlist
🎥 Visual First
Grant Sanderson · YouTube
The single best visual introduction to how neural networks actually work. Watch this before touching any code — it builds the intuition for WHY AI works the way it does. Beautifully animated, free, and deeply memorable.
Open on YouTube ↗
CS50's Introduction to AI with Python
📚 Course
Harvard University · cs50.harvard.edu/ai
Rock-solid free course from Harvard. Covers AI search, logic, ML basics with real Python projects. The quality, structure, and depth are exceptional — and it's completely free including projects and certificate.
Open Course ↗
Git & GitHub for Beginners
⚙ Tool Setup
Traversy Media · YouTube
The clearest, fastest Git tutorial. In 30–40 minutes you'll have Git configured, a repo created, and understand commit, push, pull, and branching. Everything an AI developer needs to start.
Open on YouTube ↗
Phase 0 Projects
Dev Environment Setup
Configure Python + Jupyter + Git. Write a "Hello World" notebook, commit it, and push to GitHub. Set up your first GitHub repository with a proper README.
AI Landscape Research Map
Create a visual mind map of AI subfields — ML, DL, GenAI, NLP, CV, Reinforcement Learning — with 2–3 real-world examples of each. Document it in your GitHub README.
Tip: Your GitHub profile is your first portfolio. Start maintaining it from Day 1. Commit daily — even small changes. Employers look at GitHub activity patterns.
Phase 1: Python + DSA
8–10 Weeks
Core
Beginner → Intermediate
Python is the language of AI — learn it properly, not just syntax. DSA makes you a strong problem solver and gets you through coding interviews. Combine both simultaneously: Python on weekdays, DSA on weekends.
Topics
Python — Must Know
- Variables, data types, loops, functions
- List/Dict/Set comprehensions — Pythonic style
- OOP — Classes, inheritance, dunder methods
- Error handling — try/except, custom exceptions
- Decorators, generators, context managers
- NumPy — arrays, broadcasting, vectorization
- Pandas — DataFrames, groupby, merging, cleaning
- Matplotlib + Seaborn — data visualization
DSA — Must Know
- Arrays + Strings
- Hash Maps + Sets — O(1) lookups
- Two Pointers — fundamental pattern
- Sliding Window
- Recursion + Backtracking
- Linked Lists, Stacks, Queues
- Binary Search — all variations
- Trees — BFS + DFS
- Graphs — adjacency list, BFS, DFS
- Dynamic Programming basics
Best Free Resources
Python for Everybody — Dr. Chuck
🎥 Best Beginner
Dr. Charles Severance · YouTube / Coursera (free audit)
The world's most popular Python course with 5M+ learners. Incredibly clear explanations. Covers Python basics all the way to working with data, files, APIs. Free on YouTube in full.
Open on YouTube ↗
Striver's A2Z DSA Course
📊 Best Structured DSA
TakeUForward · YouTube + strivers.io
The most structured free DSA course available. Covers every topic in order with video explanations, LeetCode problem links, and a clear progression map. Extremely trusted in the engineering community.
Open Course ↗
NeetCode 150 + Video Solutions
💡 Interview Gold
NeetCode · neetcode.io + YouTube
The gold standard curated problem list for coding interviews. 150 problems covering all patterns. Every problem has a crystal-clear video explanation. Free. NeetCode 150 is what most FAANG interviewers test.
Open NeetCode ↗
Phase 1 Projects
Titanic Data Analysis (EDA)
Full exploratory data analysis using Pandas and Matplotlib. Find survival patterns, handle missing values, visualize distributions and correlations. Document findings in a notebook with markdown explanations.
LeetCode 75 Challenge
Solve the NeetCode 75 problems (top patterns). For each solution, write a brief explanation of the pattern and time/space complexity. Push all solutions with explanations to GitHub.
CLI Task Manager
Build a command-line task manager using Python with file I/O, OOP design, and core data structures (lists, dicts, queues). Implement: add, delete, mark complete, filter, and save/load from JSON.
Phase 2: Math + Statistics for AI
6–8 Weeks
Core
Beginner → Intermediate
Don't skip math — but don't fear it. You need the intuition behind linear algebra, calculus, probability, and statistics to understand WHY ML algorithms work. Prioritize visual, animated resources over dry textbooks.
Linear Algebra (for ML)
- Vectors and vector spaces — data is always a vector
- Matrix multiplication — how neural nets compute outputs
- Transpose, inverse — essential operations
- Dot product, norms — similarity and distance
- Eigenvalues + eigenvectors — basis of PCA
- SVD — matrix factorization used in recommendations
Calculus (for ML)
- Derivatives — rate of change, the basis of learning
- Chain rule — the mathematical heart of backpropagation
- Partial derivatives — multi-variable functions in ML
- Gradient Descent — intuitively and mathematically
- Lagrange multipliers — constrained optimization
Statistics + Probability
- Mean, Median, Std Dev, Variance
- Probability distributions — Normal, Binomial, Poisson
- Bayes' Theorem — fundamental to ML thinking
- Hypothesis testing — statistical decision-making
- Confidence intervals
- Correlation vs Causation — critical for ML thinking
- MLE (Maximum Likelihood Estimation)
- Information Entropy — basis of decision trees
Best Free Resources
Essence of Linear Algebra
🎥 Best Visual
3Blue1Brown · YouTube
The most beautiful visual explanation of linear algebra ever made. Transforms abstract matrix math into geometric intuition you can actually see. Mandatory first resource before any ML course. 15 videos, completely free.
Open Playlist ↗
StatQuest with Josh Starmer
📊 Best for Stats+ML
Josh Starmer · YouTube
The best statistics educator on YouTube. Every concept is explained from absolute scratch with clear visuals and humor. Covers statistics, probability, ML algorithms, and deep learning — all completely free and extraordinarily clear.
Open Channel ↗
Khan Academy — Linear Algebra + Calculus
📚 Practice
Khan Academy · khanacademy.org/math
The most complete free math resource with interactive exercises. Use for practice problems alongside visual resources. Especially helpful for filling specific knowledge gaps in calculus and linear algebra fundamentals.
Open Khan Academy ↗
Phase 2 Projects
Gradient Descent Visualizer
Implement gradient descent from scratch in pure NumPy for a linear regression problem. Visualize the loss curve and parameter convergence across learning rates. Understand how learning rate affects training.
PCA on MNIST Digits
Apply PCA to the MNIST dataset and visualize how 784 dimensions compress to 2. Reconstruct images at different compression levels. Understand the eigenvalue contribution to explained variance.
Bayesian A/B Test Simulator
Simulate a website A/B test. Compute posterior distributions using Bayes' theorem. Make a data-driven recommendation with confidence intervals. Visualize the distributions.
Phase 3: Core Machine Learning
6–8 Weeks
Core
Intermediate
This is where AI comes alive. Learn the classic ML algorithms — understand them mathematically AND implement them. Focus on building full end-to-end pipelines, not just fitting models.
Topics
Algorithms — Must Know
- Linear Regression — regression fundamentals
- Logistic Regression — classification basics
- Decision Trees — entropy, information gain
- Random Forests — bagging ensemble
- XGBoost / LightGBM — state of the art for tabular data
- K-Means Clustering — unsupervised fundamentals
- SVMs — powerful classifier, interview topic
- Naive Bayes — fast text classification
ML Engineering — Must Know
- Train/Val/Test split — proper evaluation
- Cross-validation — robust model evaluation
- Precision, Recall, F1, ROC-AUC
- Feature Engineering — the real ML skill
- Regularization (L1/L2) — prevent overfitting
- Hyperparameter tuning — GridSearch, Optuna
- Scikit-learn Pipelines — production-ready ML
- SHAP values — model interpretability
Best Free Resources
Machine Learning Specialization — Andrew Ng
🏆 Gold Standard
DeepLearning.AI · Coursera (free audit)
The gold standard ML course. Andrew Ng's teaching clarity is unmatched anywhere in the world. Three courses covering regression, classification, neural networks, and unsupervised learning. 5M+ learners. Audit the courses for free.
Open Course ↗
Hands-On ML with Scikit-Learn & TensorFlow (notebooks)
📗 Best Book+Code
Aurélien Géron · GitHub (free notebooks)
The best practical ML book. Code-first approach with fully executable Jupyter notebooks on GitHub. Every chapter teaches theory then immediately implements it. Free notebooks with the full chapter code available on GitHub.
Open GitHub ↗
Kaggle Learn — Machine Learning Micro-Courses
💻 Interactive
Kaggle · kaggle.com/learn
Free micro-courses on ML, intermediate ML, feature engineering, and data cleaning. Interactive Jupyter environment in your browser. Learn while working on real datasets. Free certificates.
Open Kaggle Learn ↗
Phase 3 Projects
House Price Prediction Pipeline
Full ML pipeline: EDA → feature engineering → model training (Linear Regression, Random Forest, XGBoost) → hyperparameter tuning → evaluation with RMSE and R². Submit to Kaggle competition.
Customer Churn Classifier
Binary classification with class imbalance. Handle imbalance with SMOTE, engineer features, compare 3+ models, visualize ROC curves, and use SHAP to explain model decisions to stakeholders.
Movie Recommendation System
Implement collaborative filtering from scratch (user-based and item-based). Compare with Scikit-learn's SVD-based approach. Evaluate with RMSE and top-N recommendation precision.
Phase 4: Deep Learning
6–8 Weeks
Core
Intermediate → Advanced
Deep learning is the engine powering modern AI. Learn PyTorch (industry standard), CNNs, RNNs, and the Transformer architecture. Karpathy's series is legendary — prioritize it above everything else in this phase.
Topics
Neural Networks — Must Know
- MLP from scratch in NumPy — true foundations
- Backpropagation — derive it by hand once
- Activation functions — ReLU, sigmoid, softmax, GELU
- Loss functions — cross-entropy, MSE, custom
- Optimizers — SGD, Adam, learning rate schedules
- Batch Norm, Dropout, Weight Decay
- CNN architecture — Conv, pooling, image classification
- Transfer learning — fine-tune pretrained models
- RNN + LSTM — sequential data
- Transformer + Attention — the dominant architecture
PyTorch — Must Know
- Tensors + operations
- Autograd — automatic differentiation
- nn.Module — defining models
- Dataset + DataLoader — efficient data pipelines
- Training loop — forward, loss, backward, step
- GPU training — .to(device)
- Saving + loading models
- torchvision — image models
- Weights & Biases — experiment tracking
Best Free Resources
Andrej Karpathy — Neural Networks: Zero to Hero
🌟 Legendary Series
Andrej Karpathy (ex-Tesla AI, ex-OpenAI) · YouTube
Build GPT from absolute scratch, step by step. Karpathy is one of the clearest technical educators alive. This series takes you from backpropagation basics to building your own GPT — no shortcuts, deep understanding guaranteed. Free, legendary.
Open Playlist ↗
Deep Learning Specialization — Andrew Ng
🏆 Best Course
DeepLearning.AI · Coursera (free audit)
5-course deep learning specialization. Covers neural networks, CNNs, RNNs, and sequence models. Andrew Ng's explanations of backpropagation, regularization, and optimization are unmatched in clarity. Audit for free.
Open Course ↗
PyTorch Official Tutorials
📚 Official Docs
PyTorch · pytorch.org/tutorials
The official PyTorch tutorials are excellent and kept current. Cover tensors, autograd, building models, CNNs, text classification, and deployment. Use alongside Karpathy's series for the fastest path to PyTorch mastery.
Open Tutorials ↗
Phase 4 Projects
Image Classifier (CNN) from Scratch → Transfer Learning
Build a custom CNN for CIFAR-10 classification in PyTorch. Then fine-tune ResNet-50. Compare accuracy, training time, and parameters. Log everything to Weights & Biases.
Sentiment Analysis: LSTM vs Simple Transformer
Build an LSTM for IMDB sentiment analysis in PyTorch. Then implement a simple Transformer encoder. Compare both models on accuracy and training dynamics.
Build a Character-Level Language Model (Mini-GPT)
Follow Karpathy's tutorial: implement a character-level bigram model, then a full GPT with multi-head attention, positional encoding, and training loop. Train on Shakespeare text.
Phase 5: Generative AI
4–6 Weeks
Advanced
Most In-Demand Skill
Generative AI is transforming every industry and creating the highest-paying AI engineering roles. Learn how LLMs work, how to build with them, and how to ground them with your own data using RAG.
Topics
LLM Fundamentals — Must Know
- Transformer architecture deep dive
- Tokenization — BPE, SentencePiece
- Attention mechanism — self, cross, multi-head
- Prompt engineering — zero-shot, few-shot, CoT
- Embeddings — semantic representations
- Vector databases — Chroma, Pinecone, FAISS
- RAG — Retrieval Augmented Generation
- Fine-tuning with LoRA/QLoRA
- LangChain + LlamaIndex — LLM frameworks
- RLHF basics — how models are aligned
- Ollama / vLLM — local LLM inference
Free Tools & APIs
- Hugging Face Transformers — 500k+ models free
- Google Gemini API — generous free tier
- Groq — fastest free LLM inference API
- Ollama — run LLMs locally (Llama, Mistral)
- LangChain — LLM app framework
- ChromaDB — free local vector database
- Google Colab — free A100 GPU for fine-tuning
- OpenRouter — access many LLMs cheaply
Best Free Resources
LLM Course by Maxime Labonne
⭐ Best Free LLM Course
Maxime Labonne · GitHub (free, open-source)
The most comprehensive free LLM course on the internet. Covers everything from architecture to fine-tuning, quantization, RAG, evaluation, and deployment. All notebooks included, actively maintained. One of the most starred AI repos on GitHub.
Open on GitHub ↗
DeepLearning.AI Short Courses — RAG, LangChain, Fine-tuning
💡 Fastest Hands-on
DeepLearning.AI · deeplearning.ai/short-courses
20+ free 1–2 hour courses taught by the creators of LangChain, HuggingFace, Pinecone, and more. The fastest way to get hands-on with GenAI tools. Take: LangChain for LLM App Dev, Building RAG, and Fine-tuning with LoRA.
Open Courses ↗
Prompt Engineering Guide
📖 Best Prompting Resource
DAIR.AI · promptingguide.ai
The definitive free guide to prompt engineering. Covers zero-shot, few-shot, chain-of-thought, ReAct, tree of thought, and advanced techniques. Constantly updated by the research community. Essential reading for any GenAI engineer.
Open Guide ↗
Phase 5 Projects
RAG Chatbot over Your Own Documents
Build a chatbot that answers questions from a PDF/website using LangChain + ChromaDB (vector DB) + Gemini/Groq API. Add a Streamlit UI. Deploy to Hugging Face Spaces for free.
Fine-tune Llama 3 (8B) with QLoRA
Fine-tune Llama 3 on a custom dataset using QLoRA (4-bit quantization) on a free Google Colab A100. Use HuggingFace PEFT library. Evaluate before vs after fine-tuning.
Semantic Search Engine
Build a semantic search system for a text dataset using sentence-transformers for embeddings + FAISS for fast similarity search. Compare with traditional keyword BM25 search. Add Streamlit UI.
Phase 6: Agentic AI
4–6 Weeks
Frontier Skill
Advanced
Agentic AI is the fastest-growing area in AI engineering — AI that takes actions, uses tools, plans, and solves multi-step problems autonomously. This is where the highest-paying roles are being created right now.
Topics
Agent Concepts — Must Know
- ReAct pattern — Reason + Act agent loop
- Tool use / function calling — agents using APIs
- Planning and task decomposition
- Memory systems — short-term and long-term
- Agent evaluation — how to test and measure agents
- Human-in-the-loop — safety and control
- Error handling in agents — robustness
- Agent security — prompt injection defense
Frameworks — Must Know
- LangChain Agents — most widely used
- LangGraph — stateful, controllable agent graphs
- CrewAI — role-based multi-agent framework
- Tavily Search Tool — web search for agents
- OpenAI Function Calling — structured tool use
- AutoGen (Microsoft) — research-grade multi-agent
- Semantic Kernel — Microsoft production SDK
- OpenAI Assistants API — managed agent threads
Best Free Resources
AI Agents in LangGraph — DeepLearning.AI
💡 Best Free Agent Course
DeepLearning.AI + LangChain (Free)
Free course taught by the creators of LangGraph. Covers building stateful, controllable AI agents from scratch. Hands-on notebooks, real tool integrations. The fastest path from zero to working production agents.
Open Course ↗
LangGraph Documentation + Tutorials
📚 Framework Docs
LangChain · langchain-ai.github.io/langgraph
LangGraph is now the industry standard for building stateful, controllable AI agent workflows. The official docs have excellent step-by-step tutorials. LangGraph is production-ready and used by leading AI companies.
Open Docs ↗
CrewAI — Multi-Agent Framework + Examples
⚙ Multi-Agent
CrewAI · docs.crewai.com + GitHub
CrewAI makes building multi-agent teams straightforward. Role-based agents with goals, tools, and delegation. Excellent documentation with working examples of research agents, writing teams, and automated workflows.
Open Docs ↗
Phase 6 Projects
Web Research Agent
Build an agent that takes a research topic, searches the web using Tavily, reads and summarizes articles, synthesizes findings, and writes a structured research report. Implement with LangGraph for state management.
Multi-Agent Code Review System
Build a CrewAI multi-agent pipeline: Agent 1 writes code for a given spec, Agent 2 reviews it for bugs and style, Agent 3 runs tests and reports results. Full automated software development loop.
Production Customer Support Agent
Full-featured support agent with: tool use (FAQ search, order lookup), conversation memory, escalation to human, and analytics dashboard. Deploy with Streamlit. This is portfolio-worthy.
Phase 7: AI System Design + Deployment
4–5 Weeks
MLOps
Advanced
Building AI that works in notebooks is one thing. Building AI that runs reliably in production is another. Learn to design scalable AI systems, serve models via APIs, containerize with Docker, and monitor in production.
Topics
Deployment — Must Know
- FastAPI — REST API for model serving
- Docker — containerize ML apps
- Model serialization — ONNX, pickle, safetensors
- Hugging Face Spaces — free model deployment
- Streamlit Cloud — free app hosting
- Latency vs throughput — performance tradeoffs
- Caching strategies — reduce LLM API costs
- AWS/GCP basics — cloud deployment
- Kubernetes basics — container orchestration
MLOps — Must Know
- MLflow — experiment tracking, model registry
- Weights & Biases — best experiment tracker
- Model monitoring — detect drift and failures
- GitHub Actions — CI/CD for ML pipelines
- Data versioning — DVC basics
- A/B testing models — production evaluation
- Cost optimization — LLM cost management
- Prometheus + Grafana — monitoring stack
Best Free Resources
Full Stack Deep Learning Course
🏆 Best MLOps Course
FSDL · fullstackdeeplearning.com (Free)
The best free course on the full ML engineering lifecycle: tools, infrastructure, experiment tracking, deployment, and monitoring. Taught by industry practitioners who actually build production ML systems. Highly practical.
Open Course ↗
MLOps Zoomcamp
💻 Hands-on Free
DataTalks.Club · GitHub (Free)
Free end-to-end MLOps course with real projects. Covers experiment tracking with MLflow, model deployment, workflow orchestration, and monitoring. The GitHub repository has all materials and solutions.
Open GitHub ↗
FastAPI Documentation + Official Tutorial
📚 Framework Docs
Sebastián Ramírez · fastapi.tiangolo.com
FastAPI is the industry standard for serving ML models. The official docs are exceptional — clear, fast to learn, with automatic API documentation generation. The tutorial section alone is better than most paid courses.
Open Tutorial ↗
Phase 7 Projects
ML Model as REST API — Deployed
Train a model, wrap it in FastAPI with input validation, Pydantic schemas, and error handling. Containerize with Docker. Deploy to Hugging Face Spaces. Add API documentation with Swagger UI.
LLM App with Full Observability
Deploy a RAG application with MLflow tracking: log latency, token usage, retrieval scores, and error rates per request. Create a monitoring dashboard to detect quality degradation.
Automated ML CI/CD Pipeline
Set up GitHub Actions to automatically: run tests, train the model on new data, evaluate against a baseline, and deploy only if performance improves. Full automated ML workflow.
Phase 8: Portfolio + Job Preparation
4–6 Weeks (Ongoing)
Career
All Levels
Your skills only matter if employers can see them. Build a portfolio that demonstrates real engineering ability, not toy projects. A well-documented GitHub project with a live demo beats 10 undocumented notebooks.
Portfolio Strategy
GitHub Profile Essentials
- Professional README — banner, skills, current work
- 5–6 pinned repos — best projects only
- Each project README: problem, solution, demo, results
- Daily commits — green contribution graph signals activity
- Live demo links — Hugging Face Spaces or Streamlit
- Architecture diagrams in READMEs
Resume for AI Roles
- Quantify everything — accuracy %, latency ms, users
- Skills section — Python, PyTorch, LangChain, FastAPI, Docker
- Project descriptions — "Built X using Y, achieving Z"
- Keywords — LLM, RAG, fine-tuning, agentic, MLOps
- 1 page unless 5+ years experience
- ATS-friendly format — no tables, clean fonts
Best Free Resources
ML Interviews Book — Chip Huyen
📗 Best ML Interview Prep
Chip Huyen · huyenchip.com/ml-interviews-book
The definitive ML interview prep resource by a Stanford ML lecturer and author. Covers mathematical foundations, ML concepts, coding, and system design questions. Free online reading. Used by engineers at Google, Meta, OpenAI.
Open Book ↗
NeetCode 150 — Structured DSA Interview Prep
💡 DSA Interview
NeetCode · neetcode.io
The curated 150 problems covering all patterns tested in top AI company coding interviews. Every problem has a clear video solution. Use this list exclusively — it covers everything in the minimum number of problems.
Open NeetCode ↗
Awesome GitHub Profile README — Examples
⭐ Portfolio Inspiration
Community · GitHub
A curated list of outstanding GitHub profile READMEs from top engineers and data scientists. Excellent for inspiration and templates for building your own profile that stands out to recruiters.
Open GitHub ↗
Where to Apply
Job Platforms
- LinkedIn — optimize profile, apply directly
- levels.fyi — salary benchmarking
- Y Combinator Jobs — AI startup roles
- AngelList/Wellfound — startup ecosystem
- Kaggle Jobs — data science focused
- Hugging Face Jobs — AI specific roles
Target Role Titles
- AI Engineer / ML Engineer
- GenAI Engineer / LLM Engineer
- AI Application Developer
- Data Scientist (with ML focus)
- MLOps Engineer
- Prompt Engineer (entry level)
- AI Research Engineer
Week-by-Week Study Plan
~52 Weeks
2–4 Hours/Day
Beginner → Job-Ready
This is a suggested pace assuming 2–4 hours per day, 6 days a week. Move faster if you have more time. The most important thing is daily consistency — 3 hours every day beats 10 hours on weekends.
Warning: Do not skip Phases 0–2. Weak foundations cause confusion in Deep Learning and GenAI. If you rush through math, you will hit walls later and have to come back anyway.
Daily Routine Template
Each Weekday (3 hrs)
- 1 hr: Watch/read new concept (video + notes)
- 1 hr: Code or implement the concept
- 30 min: 1–2 DSA problems on LeetCode
- 30 min: Review/summarize yesterday's learning
Each Weekend (4–6 hrs)
- Review the full week's concepts (from notes)
- Build or continue a project (2–3 hrs)
- Push code to GitHub
- Write a brief learning journal entry
- Plan next week's topics
52-Week Roadmap
| Weeks | Phase | Daily Focus | Pace |
|---|---|---|---|
| 1–2 | Phase 0 | Setup dev environment. Watch 3Blue1Brown Neural Net intro (4 videos). Git + GitHub tutorial. Create first repo. | 2–3 hrs/day |
| 3–4 | Phase 0 | Harvard CS50 AI (Weeks 1–3). Explore Kaggle and Hugging Face. Understand AI landscape. | 3 hrs/day |
| 5–7 | Phase 1 — Python | Dr. Chuck Python for Everybody. Variables, loops, functions, OOP. Code daily in Jupyter. | 3 hrs/day |
| 8–10 | Phase 1 — Python | NumPy + Pandas. Titanic EDA project. Matplotlib/Seaborn. Start LeetCode easy problems. | 3–4 hrs/day |
| 11–14 | Phase 1 — DSA | Striver's A2Z DSA (Arrays → Hash Maps → Two Pointers → Binary Search). 2 problems/day. | 3–4 hrs/day |
| 15–18 | Phase 1 — DSA | Trees, Graphs, Recursion, Backtracking. NeetCode 75 list. CLI project. LeetCode medium. | 3–4 hrs/day |
| 19–21 | Phase 2 — Math | 3Blue1Brown: Essence of Linear Algebra (all 15 videos). Implement matrix ops in NumPy. | 3 hrs/day |
| 22–24 | Phase 2 — Math | 3Blue1Brown Calculus. Khan Academy practice. Implement gradient descent from scratch. | 3 hrs/day |
| 25–27 | Phase 2 — Stats | StatQuest Statistics playlist. Probability distributions. Bayes theorem. Hypothesis testing. | 3 hrs/day |
| 28–30 | Phase 3 — ML | Andrew Ng ML Specialization Course 1. Linear/Logistic Regression from scratch + Scikit-learn. | 4 hrs/day |
| 31–34 | Phase 3 — ML | Andrew Ng Course 2. Tree-based models, XGBoost, cross-validation, feature engineering. House price project. | 4 hrs/day |
| 35–37 | Phase 4 — DL | Andrew Ng DL Specialization Course 1–2. MLP from scratch. PyTorch basics. CNN image classifier. | 4 hrs/day |
| 38–40 | Phase 4 — DL | Karpathy Zero to Hero series (all videos). Build bigram model → GPT. LSTM sentiment project. | 4 hrs/day |
| 41–43 | Phase 4 — DL | Transformers + HuggingFace basics. Transfer learning. Kaggle image classification competition. | 4 hrs/day |
| 44–46 | Phase 5 — GenAI | mlabonne LLM Course (first half). DeepLearning.AI LangChain short course. Build RAG chatbot. | 3–4 hrs/day |
| 47–49 | Phase 5 — GenAI | Fine-tune Llama 3 with QLoRA (Colab). Semantic search project. Prompt engineering guide. | 3–4 hrs/day |
| 50–52 | Phase 6 — Agents | LangGraph AI Agents course. Build research agent + code review multi-agent system. | 4 hrs/day |
| 53–55 | Phase 7 — Deploy | Full Stack DL course. FastAPI + Docker. Deploy 2 projects live. MLflow experiment tracking. | 4 hrs/day |
| 56–60 | Phase 8 — Career | NeetCode 150 (2 problems/day). ML interview prep (Chip Huyen book). Portfolio polish. Apply to roles. | 4 hrs/day |
Avoiding Tutorial Hell: After watching 2 videos on any topic, stop watching and start building. If you can't implement it, you don't know it yet. That discomfort is exactly where the real learning happens.
All Projects — 24 Total
Phases 0–8
Beginner → Advanced
GitHub-Ready
Every project should be on GitHub with a clear README explaining the problem, your approach, results, and what you learned. By Phase 7–8, pick your best 5–6 to showcase as your portfolio. Click any project card to open a researched execution brief.
Phase 0: Foundation
Dev Environment Setup
Phase 0: Foundation
AI Landscape Mind Map
Phase 1: Python+DSA
Titanic EDA Analysis
Phase 1: Python+DSA
LeetCode 75 Solutions
Phase 1: Python+DSA
CLI Task Manager
Phase 2: Math+Stats
Gradient Descent Visualizer
Phase 2: Math+Stats
PCA on MNIST
Phase 2: Math+Stats
Bayesian A/B Tester
Phase 3: Core ML
House Price Prediction
Phase 3: Core ML
Customer Churn Classifier
Phase 3: Core ML
Movie Recommender
Phase 4: Deep Learning
CNN Image Classifier
Phase 4: Deep Learning
LSTM Sentiment Analysis
Phase 4: Deep Learning
Mini-GPT from Scratch
Phase 5: Generative AI
RAG Document Chatbot
Phase 5: Generative AI
Fine-tune Llama 3 (QLoRA)
Phase 5: Generative AI
Semantic Search Engine
Phase 6: Agentic AI
Web Research Agent
Phase 6: Agentic AI
Multi-Agent Code Reviewer
Phase 6: Agentic AI
Customer Support Agent
Phase 7: System Design
ML REST API (Deployed)
Phase 7: System Design
LLM App + Monitoring
Phase 8: Portfolio
Capstone AI App
Phase 8: Portfolio
Kaggle Competition Entry
Portfolio Rule: Pick your 5 best projects from Phases 3–8. Make each one exceptional: clean code, live demo link, architecture diagram in README, and a paragraph about what you learned and what you'd do differently.
Interview Preparation Guide
Start DSA from Phase 1
Ramp Up in Phase 8
8–12 Weeks Intensive
AI engineer interviews test four areas: DSA coding, ML/DL knowledge, system design, and behavioral. Start DSA in parallel with Phase 1 — consistency matters more than intensity here.
Interview Format by Company Type
| Company Type | DSA Weight | ML Theory | System Design | Key Focus |
|---|---|---|---|---|
| Big Tech (Google, Meta) | Very High | High | High | Leetcode Hard + ML System Design |
| AI Startups | Medium | High | Medium | Projects + GenAI/LLM experience |
| AI Research Labs | Low-Medium | Very High | Low | Deep ML theory + publications |
| Product Companies (AI teams) | Medium | Medium | High | MLOps + deployment experience |
| Consulting / Services | Low | Medium | Medium | Communication + practical projects |
DSA Interview Strategy
NeetCode 150 — Topic Order
- 1. Arrays & Hashing (start here)
- 2. Two Pointers
- 3. Sliding Window
- 4. Stack
- 5. Binary Search
- 6. Linked List
- 7. Trees — BFS + DFS
- 8. Heap / Priority Queue
- 9. Graphs
- 10. Dynamic Programming (last)
Problem-Solving Framework
- Read: Understand constraints and examples
- Draw: At least 2 examples on paper
- Pattern: Identify which pattern applies
- Brute force: State naive solution first
- Optimize: Improve time/space complexity
- Code: Write clean, named variables
- Test: Edge cases — empty, single, large
- Complexity: State time and space at end
Must-Know ML Interview Questions
- Explain gradient descent and its variants (SGD, Adam, RMSProp, AdaGrad)
- What is the bias-variance tradeoff? How do you detect and fix underfitting/overfitting?
- Derive backpropagation intuitively and mathematically (know the chain rule)
- Explain the Transformer architecture and self-attention mechanism in detail
- What is the difference between fine-tuning and RAG? When would you choose each?
- How do you handle class imbalance in a classification problem? (SMOTE, class weights, threshold tuning)
- Design an ML system for fraud detection / recommendation / semantic search
- What is regularization? Compare L1 (Lasso) vs L2 (Ridge) and their effects on weights
- How would you monitor an ML model in production? How do you detect and respond to data drift?
- Explain Precision, Recall, F1, ROC-AUC — when to use each and why
- What is the vanishing/exploding gradient problem? How do modern architectures address it?
- Explain the difference between BERT and GPT architecturally and in training objective
- What is RLHF? Why is it used to align LLMs? What are its limitations?
- How do LLM agents work? What are the failure modes of agentic systems?
Best Free Interview Prep Resources
| Resource | Best For | Link |
|---|---|---|
| Chip Huyen — ML Interviews Book | ML theory + system design questions | huyenchip.com ↗ |
| NeetCode 150 | DSA coding interview prep | neetcode.io ↗ |
| Striver's SDE Sheet | DSA from basics to hard | takeuforward.org ↗ |
| Pramp / Interviewing.io | Mock interviews with engineers | pramp.com ↗ |
| ML System Design (ByteByteGo) | System design for ML/AI | bytebytego.com ↗ |
| Glassdoor + Blind | Real interview questions from companies | glassdoor.com ↗ |
Best interview prep: Building real projects. If you can explain what you built, why you made each technical decision, what went wrong, and what you'd do differently — that demonstrates more than any memorized answer.
AI Engineering Tools + Tech Stack
Full Stack
Learn When Relevant
Learn each tool when it becomes relevant to your current phase. By Phase 7, you should be comfortable with all core tools. Do not try to learn everything upfront — context makes tools stick.
Languages
Core Data Libraries
ML Frameworks
GenAI + LLM Stack
Vector Databases
Deployment + MLOps
Principle: Learn the concept first, then the tool that implements it. Understanding WHY PyTorch uses autograd matters far more than memorizing its API. Tools change; concepts don't.
Learning Strategy + Anti-Patterns
Read This First
Avoid Common Mistakes
How you learn matters as much as what you learn. Most people fail not because of lack of resources but because of poor learning habits. These principles will 2–3x your learning speed.
The Core Principle
Theory vs Practice Balance
- Rule of thumb: For every 1 hour of watching/reading, spend 2 hours coding/building
- Test for understanding: Can you implement it without looking at the tutorial? If not, you don't know it yet
- Confusion is good: If you feel lost while building, that's where the deep learning happens
- Teach to learn: Explain what you just learned in writing (GitHub README, notes, blog)
Avoiding Tutorial Hell
What Tutorial Hell Looks Like
- Watching 10 videos on the same topic without coding
- Only doing tutorials with step-by-step code-along
- Never starting a project because "I'm not ready yet"
- Restarting a different course instead of finishing one
- No GitHub projects after 6 months of "learning"
What Good Learning Looks Like
- Watch 1–2 videos → immediately implement
- Build projects that break and require debugging
- Read error messages and solve them yourself first
- Commit to GitHub every single day
- One resource per topic — finish it before moving on
Effective Revision System
Spaced Repetition for AI Concepts
- Day 1: Learn new concept, code it, write notes in your own words
- Day 2: Review notes (5 min), explain concept without looking
- Day 7: Implement it again from memory in a new context
- Day 30: Use it in a project — that's when it truly sticks
- Use Anki cards for definitions, formulas, and algorithm patterns
Mental Framework for Difficult Topics
When Math Feels Hard
- Watch 3Blue1Brown before any textbook
- Ask: "What problem does this concept solve?"
- Implement the concept in NumPy before using a library
- Visualize it — plot the function, the gradient, the result
When Projects Feel Overwhelming
- Start with the simplest possible version (MVP)
- Break into tiny tasks (max 30 min each)
- Google the specific error — not "how to build a RAG system"
- Ship something broken — then fix it. Done > perfect.
The 1-year test: After following this roadmap, you should be able to: (1) Build and deploy a complete AI app, (2) Explain how Transformers work without notes, (3) Pass a coding interview with medium LeetCode problems, and (4) have 5 GitHub projects with live demos. That's what job-ready looks like.