Assignment & Project Guide: Submission, Rubric and Checklist

One repo, one story. Use exactly one repository so your progression from NumPy to GenAI is visible as one continuous artifact.

ml-dl-fundamentals/
├── README.md
├── Day-01-NumPy-Pandas/
├── Day-02-Linear-Regression/
├── Day-03-Overfitting/
├── Day-04-Fraud-Detection/
├── Day-05-Telco-Churn/
├── Day-06-Neural-Network/
├── Day-07-PyTorch/
├── Day-08-MNIST/
├── Day-09-Self-Attention/
├── Day-10-Capstone/
├── Project-Ladder/
└── .gitignore

What belongs in every substantial project

□ README.md - problem, dataset, approach, results, failure cases, learnings, run instructions □ Code/notebook - one clean entry point; requirements.txt where a reproducible environment is needed □ results/ - saved plots and result artifacts with descriptive filenames □ No large datasets. No secrets. No large model weights unless intentionally small and required

.gitignore baseline

data/
*.pt
__pycache__/
.ipynb_checkpoints/

Root README table (fill one row per day)

Use for Mini Project 1 (Telco), Mini Project 2 (MNIST) and the capstone. This is a self-check, not an exam. Fix the lowest-scoring row first.

Universal README template

# Project name
## Problem
## Dataset
## Approach / Architecture
## Results (table + plots)
## Metrics
## Failure cases
## What I learned
## How to run
## Future improvements

Capstone README template

# <Project title>
One-sentence summary.

## Problem
## Dataset            (source, size/splits)
## Model / Approach   (architecture and shapes)
## Training setup     | Optimizer | LR | Batch | Epochs | Regularization |
## Results            | Model | Val acc | Test acc |   + results/loss_curves.png
                                                       + results/confusion_matrix.pn
                                                           g
## Failure cases
## What I learned
## Things I had to look up
## How to run        pip install -r requirements.txt  /  python train.py
## Future improvements

Five reflection questions after each major project

1. What problem did you solve?
2. Why did you choose this model?
3. Which metric mattered, and why?
4. What failed?
5. What would you improve next?

Cause one controlled failure per major day so the concept becomes memorable. Write one sentence per step.

Use this before calling the course complete

□ Days 2–3: both cosine similarity and Titanic cleaning done □ Day 4: scratch gradient descent on the book's generated data, compared with scikit-learn □ Day 5: degrees 1-15 evaluated; Ridge + Lasso in the 5-fold CV comparison (4 models) □ Day 6: stratified fraud evaluation with precision, recall, F1, ROC-AUC and confusion matrices □ Days 7–8: k-NN, k-Means, PCA, embeddings and the Telco Churn project □ Days 9–10: gradient check + hidden-size, learning-rate and activation experiments □ Day 11: all six optimizer runs (SGD/Adam x 0.1/0.01/0.001), plotted, state_dict saved □ Day 12: 55k/5k/10k split, MLP, dropout, CNN, early stopping, test evaluation, confusion matrix, at least 5 mistakes shown □ Day 13: Q/K/V, sqrt(d_k) scaling, softmax, shapes, row sums, causal mask with future zeros, heatmap, Transformer block diagram □ Days 14–15: one capstone path rebuilt from a blank file and published □ Part II (optional ladder): RAG cites sources and says “I don't know”; tools validate arguments; agents have a step limit and logs; MCP tool runs as a server; the production API has FastAPI, Docker, an eval set and cost/latency numbers □ Dataset filenames are consistent (including the Telco rename note); no large data or secrets in GitHub □ Optional Level 2 work was never used to satisfy a required Level 1 item

Progress tracker

Tick each box when done.

Remember The book teaches; this guide makes you build, verify, reflect and submit. Do Level 1 every day, and add Level 2 as you grow. Finished and simple beats perfect and unfinished.