AI from scratch

Plan · 6 to 8 months

From zero to interview-ready

A detailed path through the eight modules — learn deeply, build projects, write everything down, and drill the questions that actually show up in ML / GenAI interviews.

This roadmap assumes ~12–15 hours a week (stretch to ~20 if you can). The rule that makes it stick: every major concept becomes a short write-up on this site plus one runnable notebook. If you skip writing, you will not retain it at interview depth.

By month 6–8 you should be able to: explain any core idea in plain English, implement small experiments (sklearn + PyTorch), answer ML / DL / LLM questions with tradeoffs, walk through 1–2 portfolio projects, and handle a light GenAI system-design round.

On this page

Each concept below has two links when available: Lesson → the module section that teaches it, and Drill → interview questions. “Soon” means the module isn’t published yet — drill from the interview bank meanwhile.

Concept index

Jump straight to a concept. Live lessons link into Modules 1–3; later topics link to the interview bank until those modules ship.

Foundations · live

Programming vs learning (the mental flip) LessonDrill
Supervised / unsupervised / reinforcement learning LessonDrill
Train / validation / test split & leakage LessonDrill
Overfitting, underfitting, bias–variance LessonDrill
Linear & logistic regression LessonDrill
Decision trees, bagging, boosting LessonDrill
Tabular vs unstructured — when to use what LessonDrill

Deep learning · live + extend

Neurons, layers, activations LessonDrill
Loss, gradients, backpropagation LessonDrill
Why non-linearity matters (MLP vs logistic) LessonDrill
Deep learning → GenAI bridge LessonDrill
CNNs (conv, pool, vision) LessonDrill
Regularization: dropout, batch norm, learning rate LessonDrill
RNNs → why attention won Module soonDrill

Transformers & GenAI · live

Tokenization & embeddings LessonDrill
Self-attention (Q, K, V) LessonDrill
Transformer block & GPT vs BERT LessonDrill
Pretraining, SFT, RLHF / DPO LessonDrill
RAG, hallucinations, LLM eval LessonDrill
Temperature, top-p, prompting LessonDrill

Data, MLOps, systems · live

Data leakage, splits, imbalance LessonDrill
Feature importance / explainability LessonDrill
Serving, monitoring, drift LessonDrill
RAG bot / search / moderation designs LessonDrill

Weekly template

Repeat every week:

Learn     40%   one concept deeply
Build     30%   notebook or project slice
Write     15%   publish on this site
Drill     15%   interview Q&A out loud, timed

Drill bank: Interview questions — answer five per week in writing.

Phase 0 — Foundations Weeks 1–3 · Modules 1–2

You may already have most of this live. Treat it as a refresh until you can teach it.

The mental flip — rules vs learning from examples LessonDrill
Supervised / unsupervised / RL (+ LLM analogies) LessonDrill
Train / val / test, leakage, metrics LessonDrill
Overfitting & bias–variance (taxi experiment) LessonDrill
Linear & logistic regression LessonDrill
Trees, random forests, gradient boosting LessonDrill
When classic ML beats deep learning LessonDrill

Checkpoint: explain overfitting and bagging vs boosting on a whiteboard in under five minutes.

Module 1 · Module 2 · Foundation Qs · Algorithm Qs

Phase 1 — Deep Learning core Weeks 4–8 · Module 3

Week 4 — Neurons, layers, activations LessonDrill
Week 5 — Loss, gradient descent, backprop LessonDrill
Week 5 — Non-linear boundaries (circles experiment) LessonDrill
Week 6 — CNNs (Fashion-MNIST / CIFAR in PyTorch) LessonDrill
Week 7 — Dropout, batch norm, learning rates LessonDrill
Week 8 — RNNs (light) → why attention replaced them Write-up soonDrill
How deep learning unlocks GenAI LessonDrill

Project A: image or tabular classifier with clean splits, metrics, and a short failure write-up.

Checkpoint: derive backprop for a 2-layer net; explain vanishing gradients and residuals.

Tools: Python, NumPy, pandas, matplotlib, scikit-learn, PyTorch.

Module 3 · Deep learning Qs

Phase 2 — NLP → Transformers → LLMs Weeks 9–16 · Modules 4–5

Weeks 9–11 — NLP & tokenization

Bag-of-words / TF-IDF and their limits LessonDrill
Embeddings, tokens, BPE, vocab tradeoffs LessonDrill
Next-token prediction as the core objective LessonDrill

Weeks 12–14 — Transformers (Module 4)

Self-attention (Q, K, V) — tiny example by hand LessonDrill
Multi-head attention & positional encoding LessonDrill
Encoder / decoder / GPT-style vs BERT-style LessonDrill

Weeks 15–16 — LLMs in practice (Module 5)

Pretraining vs SFT vs RLHF / DPO LessonDrill
Prompting, few-shot, chain-of-thought LessonDrill
RAG pipeline & failure modes LessonDrill
Hallucinations & evaluation LessonDrill

Project B (highest interview signal): RAG over your own notes/blog — chunking, embeddings, retrieval, citations, and a 20-question eval set.

Checkpoint: whiteboard attention, next-token training, and RAG in one sitting.

LLM / GenAI interview Qs

Phase 3 — Data & MLOps Weeks 17–22 · Modules 6–7

Data & features (Module 6)

Missing values, encoding, scaling, leakage LessonDrill
Time-based splits, imbalance, label noise LessonDrill
Feature importance / SHAP intuition LessonDrill

MLOps & deployment (Module 7)

Serving (API) & experiment tracking LessonDrill
Monitoring: drift, latency, cost LessonDrill
LLM ops: prompt versions, eval, guardrails LessonDrill

Project C: deploy Project A or B; document how you’d know it’s rotting in production.

Checkpoint: five-minute answer to “How do you put this in production and detect decay?”

Data Qs · MLOps Qs

Phase 4 — Systems & interview polish Weeks 23–32 · Module 8

If you only have six months, compress this into weeks 23–26 and keep drilling.

System designs to practice

RAG customer-support bot LessonDrill
Semantic / hybrid search LessonDrill
Content moderation pipeline LessonDrill
Fine-tune vs prompt vs RAG decision LessonDrill
Recommendation feed LessonDrill
Document Q&A over private PDFs LessonDrill

Weekly interview mix

ML theory     5–10 Qs from the interview bank
Coding        Easy/medium Python + pandas; occasional “implement X”
Deep learning Tradeoffs: CNN vs Transformer, LR, overfitting signs
GenAI         Tokens, context, temperature, RAG vs fine-tune, eval
Behavioral    Your “learning in public” story — this site is the proof

System design Qs · Full interview bank

Portfolio projects

Hands-on work is tracked on the Projects page — real Finance & Energy systems (not curriculum demos).

Open the projects tracker →

Math you actually need

Do this alongside Phases 1–2 (2–3 hrs/week) — not as a two-month prelude.

Linear algebra — vectors, matrices, dot products Related lessonAttention drill
Probability — expectation, conditionals, Bayes Related lessonDrill
Calculus — derivatives & chain rule (backprop) LessonDrill
Stats — mean/variance, distributions (high level) Related lessonDrill

6-month vs 8-month

6 months (aggressive)
  Phases 0→1→2→3, then ~4 weeks of interviews
  Projects A + B required; C light
  Math: intuition-heavy

8 months (recommended for “everything in detail”)
  Full Phase 4 with more reps
  Projects A + B + C polished
  Light linear algebra / probability refresh
  Read one transformer implementation carefully

Mapped to the curriculum

Module 1 — ML Foundations · Phase 0 · High OpenDrill
Module 2 — Classic Algorithms · Phase 0 · High OpenDrill
Module 3 — Deep Learning · Phase 1 · High OpenDrill
Module 4 — Transformers & LLMs · Phase 2 · Highest OpenDrill
Module 5 — Generative AI · Phase 2 · Highest OpenDrill
Module 6 — Data & Features · Phase 3 · Medium–high OpenDrill
Module 7 — MLOps & Deployment · Phase 3 · Medium OpenDrill
Module 8 — System Design & Interviews · Phase 4 · Highest OpenDrill
Right now: work Phase 0–1 concepts with live lesson links, start a Module 4 outline (attention → transformer → GPT training loop), get comfortable with PyTorch on one small tutorial, and answer five questions a week from the interview bank.