AI Master Roadmap
From math foundations to LLMs and MLOps — phased learning path with projects and interview milestones.
Interview tip Pick ONE phase to go deep this month. Ship a project per phase — hiring managers care about demos, not certificates.
① Roadmap overview — five phases
AI learning path
Math & Stats→Classical ML→Deep Learning→LLMs & Apps→MLOps
| Phase | Duration | Outcome | Project |
|---|---|---|---|
| 0 — Foundations | 4–6 weeks | Linear algebra, probability, Python | NumPy matrix ops + EDA notebook |
| 1 — Classical ML | 6–8 weeks | Supervised/unsupervised, sklearn | Kaggle tabular competition top 50% |
| 2 — Deep Learning | 8–10 weeks | CNNs, RNNs, training loops | Image classifier or seq2seq model |
| 3 — LLMs & GenAI | 6–8 weeks | Transformers, RAG, agents, fine-tuning | RAG chatbot over your docs |
| 4 — MLOps | 4–6 weeks | Deploy, monitor, CI/CD for models | API endpoint + drift monitoring |
② Phase 0 — Math & programming foundations
- Linear algebra: vectors, matrices, dot product, eigenvalues (PCA intuition)
- Calculus: derivatives, chain rule (backprop intuition)
- Probability: Bayes, distributions, expectation, variance
- Statistics: hypothesis testing, confidence intervals
- Python: NumPy, pandas, matplotlib; comfortable with Jupyter
- Git basics: commit, branch, push — all projects in GitHub
Resources: 3Blue1Brown (Essence of LA), StatQuest (YouTube), Khan Academy, "Mathematics for ML" (Deisenroth). Don't get stuck — 80% intuition, move on when you can follow ML lectures.
③ Phase 1 — Classical machine learning
| Topic | Key concepts | Interview must-know |
|---|---|---|
| Supervised | Regression, classification, train/val/test split | Bias-variance, cross-validation |
| Algorithms | Linear/logistic regression, trees, RF, XGBoost, SVM | When to use which; interpretability |
| Unsupervised | K-means, PCA, anomaly detection | Choosing k; PCA for dimensionality reduction |
| Evaluation | Accuracy, precision, recall, F1, ROC-AUC | Imbalanced classes → do not trust accuracy |
| Feature eng | Scaling, encoding, missing values, leakage | Target leakage is a common interview trap |
Project: End-to-end sklearn pipeline on a real dataset — document EDA, feature choices, and why you picked the final model. Deploy nothing yet; focus on methodology.
④ Phase 2 — Deep learning
Neural network basics — layers, activations, loss, optimizer
Backpropagation intuition — chain rule through the graph
CNNs — conv, pool, ResNet; image classification
RNNs/LSTMs — sequences, vanishing gradient; time-series/NLP pre-transformer
Training — batch size, learning rate schedule, regularization (dropout, weight decay)
Frameworks — PyTorch preferred for research/jobs; TensorFlow for some enterprise
Resources: fast.ai Practical Deep Learning, Andrew Ng Deep Learning Specialization, PyTorch tutorials. Project: Train a CNN on CIFAR-10 or fine-tune a small model — log metrics in TensorBoard or W&B.
⑤ Phase 3 — LLMs & generative AI
| Topic | What to learn | Hands-on |
|---|---|---|
| Transformers | Self-attention, positional encoding, encoder-decoder | Read "Attention Is All You Need" summary |
| Pre-trained models | GPT, BERT, T5 — pretrain vs fine-tune vs prompt | Hugging Face model hub experiments |
| Prompt engineering | System prompts, few-shot, chain-of-thought | Build eval set of 20 prompts |
| RAG | Embeddings, chunking, vector DB, retrieval + generation | Chatbot over PDF/wiki |
| Fine-tuning | LoRA, RLHF concept, when to fine-tune vs RAG | Fine-tune small model on domain data |
| Agents | Tool use, ReAct loop, guardrails | Agent with 2 tools (search + calculator) |
Use cheat sheets on this site: RAG, First LLM Integration, Agents Intro, Prompt Engineering.
⑥ Phase 4 — MLOps & production
- Model serving: FastAPI + ONNX/TorchServe/vLLM for LLMs
- Containerization: Docker basics; deploy to AWS/GCP/Fly.io
- Monitoring: latency, error rate, data drift, model drift
- CI/CD: retrain pipeline triggered on new data
- Cost: token usage, GPU hours, cache embeddings
- Security: API keys in secrets manager, input validation, PII in logs
Minimal ML production stack
Client / App
API Gateway + Auth
Model Server (GPU/CPU)
Vector DB + Feature Store
Logging / Metrics (Prometheus, Grafana)
⑦ Role-specific tracks
| Target role | Emphasize | De-emphasize |
|---|---|---|
| ML Engineer | Training pipelines, distributed training, model optimization | Heavy frontend |
| Data Scientist | EDA, experimentation, causal inference, storytelling | CUDA kernel tuning |
| AI App Developer | LLM APIs, RAG, agents, product UX | Proof of convergence theorems |
| Research Scientist | Paper reading, novel architectures, benchmarks | CRUD app building |
Most industry roles in 2025–2026 are AI application + ML engineering hybrid — strong Python, LLM integration, and evaluation beat pure theory.
⑧ 12-week sprint plan (template)
| Weeks | Focus | Deliverable |
|---|---|---|
| 1–2 | Math refresh + Python | GitHub repo with NumPy exercises |
| 3–4 | Classical ML + Kaggle | One competition submission + write-up |
| 5–6 | PyTorch + CNN project | Image model with training curves logged |
| 7–8 | Transformers + Hugging Face | Fine-tune or zero-shot eval notebook |
| 9–10 | RAG application | Deployed chatbot (even on free tier) |
| 11 | Mock ML interviews | 5 timed Q&A sessions recorded |
| 12 | Portfolio polish | README, demo video, LinkedIn post |
Adjust pace: 10–15 hrs/week realistic alongside a job. Double timeline if <8 hrs/week.
⑨ Interview prep milestones
- Explain bias-variance without notes in under 2 minutes
- Whiteboard: train/val/test split and k-fold CV
- Describe attention mechanism at high level
- Compare RAG vs fine-tuning with a concrete example
- Walk through one project end-to-end (problem → data → model → metric → result)
- Answer "what would you do with imbalanced classes?" with 3 techniques
- Discuss one failure: wrong metric, data leakage, or overfitting you fixed
⑩ Revision checklist
- Identified current phase and next project
- Balanced theory with shipped projects on GitHub
- Know which role track to optimize for
- Have 12-week plan with weekly deliverables
- Prepared 2-minute project stories for interviews
- Connected roadmap to site cheat sheets (RAG, agents, etc.)