/ ML, DL & GenAI / Phase 3 Phase 3 Days 45-64
Unsupervised Learning, Recommenders, and Retrieval Foundations PCA, clustering, anomaly detection, recommenders, embeddings, vector search, hybrid retrieval, labeling loops
Phase Goal Build useful systems when labels are missing: discover structure, rank items, retrieve information, and evaluate fuzzy outputs with human review loops — before any LLM-based RAG.
Day 45: Unsupervised Learning Map
Evaluation & Failure Modes When unsupervised work is genuinely useful and when it is a research detour the business does not need. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 46: PCA, Eigenvectors, and SVD
Evaluation & Failure Modes Reading projection plots without overclaiming. The "PCA cluster" mirage that fools many notebooks. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 47: Manifold Learning: t-SNE and UMAP
Evaluation & Failure Modes Visual overclaiming. The rules about what you may and may not conclude from a t-SNE plot. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 48: K-Means From Scratch
Evaluation & Failure Modes Cluster stability across seeds. Choosing k with the elbow rule, silhouette, and gap statistic — and why none of them is definitive. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 49: Hierarchical Clustering
Evaluation & Failure Modes Granularity decisions: who chooses the cut, and how to defend it to a stakeholder. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 50: DBSCAN and HDBSCAN
Evaluation & Failure Modes Density sensitivity: when DBSCAN cannot find clusters of varying density and HDBSCAN saves you. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 51: Gaussian Mixture Models and EM
Evaluation & Failure Modes Identifiability problems. Soft-cluster quality vs hard-cluster quality. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 52: Anomaly Detection With Statistics
Evaluation & Failure Modes False-alert fatigue. Why most anomaly systems die from operator burnout, not algorithm choice. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 53: Isolation Forest, LOF, One-Class Models
Evaluation & Failure Modes Alert calibration and how to put an SLA on a fuzzy detector. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 54: Association Rules and Market Basket Mining
Evaluation & Failure Modes Rule usefulness. Why a high-lift rule can still be commercial garbage. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 55: Recommendation Systems Overview
Evaluation & Failure Modes When ranking metrics disagree with business metrics, which one wins. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 56: Matrix Factorization and Embeddings
Evaluation & Failure Modes Embedding behaviour: popularity bias, dimension collapse, and the cold-start cliff. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 57: Ranking Metrics and Offline Evaluation
Evaluation & Failure Modes Offline-online mismatch: why offline-best is often online-worst. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 58: Content-Based and Hybrid Recommenders
Evaluation & Failure Modes Cold-start performance: which hybrid actually rescues new items vs new users. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 59: Embeddings and Vector Search Foundations
Evaluation & Failure Modes Retrieval relevance, judged by humans on a small gold set — the cheapest reliable signal. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 60: ANN Search: HNSW, IVF, Product Quantization
Evaluation & Failure Modes Index quality regressions as a vector store grows. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 61: Hybrid Search and Reranking
Evaluation & Failure Modes Search failure analysis: the manual review workflow that finds the failure mode no metric catches. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 62: Topic Modeling and Label Discovery
Evaluation & Failure Modes Topic validity. The "we have ten topics" trap when half are garbage. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 63: Active Learning and Human Labeling Loops
Evaluation & Failure Modes Label quality as a first-class metric. The rubric is the model. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Day 64: Capstone: Discovery, Retrieval, or Recommender System Final lab: build a segmentation, anomaly, recommender, topic, or retrieval system. Prove usefulness with examples and a review workflow. Use only tools already taught: explain an early library call fully, then rebuild its mechanism once prerequisites are ready; offer an executable small-data or CPU route Add shape, dtype, and range assertions so silent bugs become loud bugs Compare with a taught reference under matched data, objective, dtype, and justified tolerance; do not demand identical stochastic runs
Evaluation & Failure Modes System usefulness, judged the way a product manager would judge it. Inspect individual errors, slices, prompts, or traces — never trust a single average score Catalogue leakage, overfitting, bias, hallucination, latency, cost, or operational risks for this technique Write the next experiment from the failure mode you observed Close the notebook and reconstruct one mechanism; log help used and schedule a delayed changed-task check, initially around 1/3/7/14/30 days Keep ordinary review within 20 minutes; if weak prerequisites accumulate, pause new material and repair one dependency Record two competing explanations and the smallest experiment that distinguishes them; distinguish a negative result from an implementation failure
Capstone: Discovery, Retrieval, or Recommender System Build a segmentation, anomaly, recommender, topic, or retrieval system and prove usefulness with examples, stability checks, and a human review workflow.
After this phase, you'll be able to:
Use similarity spaces and dimensionality reduction responsibly Cluster, profile, and detect anomalies in unlabeled data Build recommender and retrieval baselines Evaluate discovery systems with examples and ranking metrics You can build discovery, recommendation, and retrieval systems before touching LLM-based RAG. Gate: explain without notes, build one independent artifact, diagnose a deliberate failure, and repeat a changed task after a delay. Record help and repair missing prerequisites.