🗓️ 🗓️ All 90 Days Reference — AI/LLM Roadmap

Every day has a specific deliverable: build something from scratch, then verify it against the library version. Mark Done only when you can explain the mechanism, not just call the function.

---

Phase 1 — Python & Math Foundations (Days 1–15)

Day	Topic	Daily Milestone
1	NumPy vectorization	Replace 3 loop-based computations with vectorized NumPy. Benchmark speedup.
2	NumPy broadcasting	Normalize a dataset using broadcasting. Explain axis=0 vs axis=1.
3	NumPy linear algebra ops	Matrix multiply, dot product, norm. Implement cosine similarity from scratch.
4	Pandas basics	Load CSV, inspect, filter, groupby on a real dataset.
5	Pandas missing data	Handle missing values 3 ways (drop, median impute, flag). Compare results.
6	Pandas encoding	One-hot encode categoricals. Prep a train/test split with no leakage.
7	Vectors and dot products	Compute dot product by hand and with NumPy. Explain geometric meaning.
8	Matrices as transformations	Visualize a matrix scaling/rotating vectors. Explain what Wx does.
9	Eigenvalues and eigenvectors	Implement PCA from scratch using eigendecomposition. Compare to sklearn PCA.
10	Matrix rank, determinant, inverse	Compute all 3 for a 3x3 matrix by hand and with NumPy.
11	Derivatives and gradients	Implement numerical derivative. Compare to analytical derivative for x^2.
12	Chain rule	Derive d/dx of a composed function by hand. Verify with PyTorch autograd.
13	Gradient descent	Implement gradient descent to minimize (x-3)^2. Plot convergence.
14	Probability distributions + Bayes	Implement Bayes theorem on a concrete example (medical test scenario).
15	Phase 1 Capstone	Linear regression from scratch with gradient descent. Match sklearn's weights. Plot loss curve.

---

Phase 2 — Classical Machine Learning (Days 16–35)

Day	Topic	Daily Milestone
16	Linear regression with sklearn	Fit on a real dataset with a Pipeline + StandardScaler.
17	Logistic regression	Fit a binary classifier. Inspect predict_proba outputs.
18	L1 vs L2 regularization	Train Ridge and Lasso on the same data. Compare which features Lasso zeros out.
19	Bias-variance tradeoff	Train models of increasing complexity. Plot train vs test error. Identify overfit point.
20	Cross-validation	Replace single train/test split with 5-fold CV. Report mean +/- std.
21	Classification metrics	Compute precision, recall, F1, confusion matrix on an imbalanced dataset.
22	Decision trees	Train and visualize a decision tree. Explain a split using Gini impurity.
23	Random Forest (bagging)	Train Random Forest. Compare variance to a single decision tree.
24	Gradient Boosting	Train XGBoost with early stopping. Compare to Random Forest.
25	Feature importance	Extract and plot feature importances from your best tree model.
26	SVM	Train SVM with RBF kernel. Explain the kernel trick conceptually.
27	KNN	Train KNN classifier. Demonstrate performance degradation in high dimensions.
28	Curse of dimensionality	Simulate: compute average distance between random points as dimensions increase.
29	K-Means clustering	Implement elbow method. Cluster a dataset and visualize.
30	DBSCAN	Compare DBSCAN to K-Means on a non-spherical dataset.
31	PCA for visualization	Reduce a high-dim dataset to 2D with PCA. Plot and interpret.
32	t-SNE	Compare PCA vs t-SNE visualization on the same dataset.
33	Feature engineering pipeline	Build a ColumnTransformer handling numeric + categorical features.
34	Hyperparameter tuning	GridSearchCV on your best model. Report best params and CV score.
35	Phase 2 Capstone	End-to-end ML pipeline: EDA, 3 model families, tuning, final eval, feature importance, report.

---

Phase 3 — Deep Learning Foundations (Days 36–55)

Day	Topic	Daily Milestone
36	Neural network forward pass	Implement forward pass (2-layer) in pure NumPy.
37	Activation functions	Implement sigmoid, ReLU, softmax. Plot each. Explain why non-linearity matters.
38	Backprop derivation	Derive gradients for a 2-layer network by hand on paper.
39	Backprop implementation	Implement backward pass in NumPy. Train on a toy classification dataset.
40	Chain rule in backprop	Trace gradient flow through 3 layers by hand. Verify against autograd.
41	PyTorch autograd	Use torch.autograd to compute gradients. Compare to your NumPy implementation.
42	Vanishing/exploding gradients	Train a deep network with sigmoid vs ReLU. Compare gradient magnitudes.
43	PyTorch nn.Module	Build an MLP using nn.Module. Train on MNIST.
44	PyTorch training loop	Write the full loop: zero_grad, forward, loss, backward, step. Train to convergence.
45	Optimizers: SGD vs Adam	Train the same model with SGD and Adam. Compare convergence speed.
46	DataLoaders and batching	Use DataLoader with batching + shuffling. Explain why batching matters.
47	model.eval() and dropout	Add Dropout. Show different behavior in train() vs eval() mode.
48	CNN: convolution operation	Implement a conv layer forward pass conceptually. Explain parameter sharing.
49	CNN: build and train	Build a CNN in PyTorch. Train on CIFAR-10 or similar.
50	CNN: pooling and architecture	Add pooling layers. Compare model size vs a fully-connected equivalent.
51	Transfer learning	Fine-tune a pretrained ResNet. Compare to your from-scratch CNN.
52	RNN basics	Build a simple RNN. Explain the hidden state mechanism.
53	LSTM/GRU	Replace RNN with LSTM. Explain why it handles long sequences better.
54	RNN limitations	Document why RNNs can't parallelize across sequence length (sets up Phase 4).
55	Phase 3 Capstone	CNN image classifier + LSTM char-level language model with text generation at 3 temperatures.

---

Phase 4 — NLP & Transformers (Days 56–70)

Day	Topic	Daily Milestone
56	Tokenization	Tokenize text with a BPE tokenizer. Inspect subword splits on rare words.
57	BPE algorithm	Implement a simplified BPE merge algorithm conceptually on a tiny corpus.
58	Embeddings	Use nn.Embedding. Explore word similarity using cosine distance on embeddings.
59	Self-attention derivation	Derive the QKV attention formula on paper. Explain each matrix's role.
60	Self-attention implementation	Implement scaled dot-product attention from scratch in PyTorch.
61	Causal masking	Implement causal masking. Verify: changing a future token doesn't affect past outputs.
62	Attention scaling	Explain why sqrt(d_k) scaling matters. Show softmax saturation without it.
63	Multi-head attention	Implement multi-head attention from scratch (no nn.MultiheadAttention).
64	Residual connections + LayerNorm	Add both to your transformer block. Explain why they help gradient flow.
65	Positional encoding	Implement sinusoidal positional encoding. Explain why attention needs it.
66	Full Transformer block	Assemble: attention + residual + FFN + residual + norm. Stack N blocks.
67	BERT architecture	Load a pretrained BERT. Explain encoder-only + bidirectional + MLM pretraining.
68	GPT architecture	Load GPT-2. Generate text. Explain decoder-only + causal + next-token pretraining.
69	BERT vs GPT decision	Write up: when would you use an encoder vs decoder architecture? With examples.
70	Phase 4 Capstone	GPT-style transformer from scratch (no nn.Transformer) + fine-tuned BERT on a real task.

---

Phase 5 — LLMs & Applied GenAI Engineering (Days 71–90)

Day	Topic	Daily Milestone
71	Temperature sampling	Implement temperature scaling. Generate text at temp 0.1, 1.0, 2.0. Compare.
72	Top-k and top-p sampling	Implement both from scratch. Explain when to use each.
73	Prompt engineering patterns	Write few-shot and chain-of-thought prompts. Compare output quality to zero-shot.
74	Embeddings for retrieval	Embed a small document set. Compute cosine similarity for a query.
75	FAISS vector search	Build a FAISS index. Run nearest-neighbor search. Explain HNSW conceptually.
76	Vector DB comparison	Compare FAISS vs Chroma vs Pinecone tradeoffs (local vs managed, scale).
77	RAG: chunking strategy	Chunk a document 3 ways (size/overlap). Compare retrieval quality qualitatively.
78	RAG: full pipeline	Build index -> retrieve -> generate pipeline with LangChain or from scratch.
79	RAG: source citations	Modify your pipeline to always return source documents with the answer.
80	RAG: hybrid search	Add BM25 keyword search alongside vector search. Compare results.
81	RAG: re-ranking	Add a cross-encoder re-ranker. Measure precision improvement on top-K.
82	RAG: failure modes	Test your RAG on a question with no good answer in your docs. Does it say "I don't know"?
83	LoRA concept and config	Set up a LoraConfig. Print trainable params %. Explain the low-rank math.
84	LoRA fine-tuning	Fine-tune a small model with LoRA on a custom dataset.
85	QLoRA and quantization	Load a model in 4-bit. Fine-tune with QLoRA. Compare memory usage to full fine-tune.
86	Fine-tune vs RAG decision	Write a decision framework: when to fine-tune, when to RAG, when to just prompt.
87	Function calling / tool use	Implement a function-calling agent with 2 tools. Test multi-step tool use.
88	ReAct agent pattern	Build a ReAct loop: thought -> action -> observation -> repeat until answer.
89	Evaluation harness	Build a 20-question golden eval set. Run faithfulness + relevance metrics.
90	Phase 5 Capstone	Production RAG app: ingestion + retrieval + generation + evals + FastAPI + UI + README with limitations.

---

Daily ritual

1. Open this table. Find today's row.
2. Read the concept explanation in the phase page.
3. Implement the "from scratch" version first — resist reaching for the library immediately.
4. Verify correctness against the library/pretrained version.
5. Log in tracker: tick Built from scratch and Used the library version.
6. Write one Key insight in your own words — not copied from the explanation.
7. Mark Done only if you could explain today's mechanism to someone else without notes.