Day 15 · What Comes Next
Suggested pace: 1–1.5 hrs learn · 2–3 hrs code · rest to debug and document What you should know first ☐ Day 13: attention and the Transformer idea ☐ Day 7: embeddings and cosine similarity ☐ Day 6: precision and recall
Suggested pace: 1–1.5 hrs learn · 2–3 hrs code · rest to debug and document
What you should know first ☐ Day 13: attention and the Transformer idea ☐ Day 7: embeddings and cosine similarity ☐ Day 6: precision and recall If one of these feels shaky, spend 15 minutes revisiting it before starting. That is normal and saves time later.
Goal of today Connect everything you learned to modern GenAI (embeddings, vector search, RAG, agents, tools and MCP), then use the study toolkit to revise and plan your next projects.
You will be able to:
- Explain how Transformers lead to LLMs
- Describe the path from embeddings to RAG and agents
- Choose a GenAI project from the project ladder
- Revise with the master checkpoint bank
- Plan what to learn next
PART J — WHAT COMES NEXT (the GenAI roadmap explained simply)
LLMs → GenAI → Prompt Engineering → Embeddings → Vector DBs → RAG → Fine-tuning
→ AI Agents → Tool Calling → MCP → LangChain → LangGraph → AI Evaluation
→ AI Security → FastAPI → Docker → LLMOpsStage What it is (simple) Builds on what you learned LLMs Huge decoder-only Transformers Day 13 trained to predict the next token
| Stage | What it is (simple) | Builds on what you learned |
|---|---|---|
| GenAI | Models that generate text/images/code/audio | Day 13 |
| Prompt Engineering | Writing instructions/examples to steer an LLM | — |
| Embeddings | Meaning as vectors | Days 2–3, 5 (cosine similarity) |
| Vector Databases | Store embeddings, find nearest neighbors fast (FAISS, Chroma, pgvector, Pinecone) | Days 7–8 (k-NN idea) |
| RAG | Retrieve relevant docs by embedding similarity → put in prompt → grounded answer | Embeddings |
| Fine-tuning | Continue training a model on your data (LoRA/QLoRA) | Day 11 (PyTorch), Day 4 (GD) |
| AI Agents | LLM in a loop: think → act → observe → repeat | LLM + tools |
| Tool Calling | LLM outputs structured function calls your code executes | Agents |
| MCP | Standard protocol connecting LLMs to tools & data sources | Tool calling |
| LangChain | Framework to chain LLM steps, retrievers, tools | RAG/Agents |
| LangGraph | Build stateful, multi-step agent workflows as graphs | Agents |
| AI Evaluation | Measure quality: test sets, RAGAS, LLM-as-judge | Day 6 (metrics) |
| AI Security | Prompt injection, data leaks, guardrails | — |
| FastAPI | Serve models/apps via HTTP APIs | Backend skills |
| Docker | Package and ship reproducibly | DevOps |
| LLMOps | Monitoring, tracing, cost/latency, versioning in production | Production |
How to study each next topic (repeat this recipe)
- Understand — one video/article, write a 5-line summary
- Build — a tiny project in <1 day
- Break — test failure cases
- Document — README + short post
- Repeat — link to the previous topic (e.g. RAG uses embeddings + vector DB + LLM)
✓ Days 14–15 Deliverables
□ Capstone code, clean and runnable □ README with results & plots □ Explanation of the 6 parts (in README) □ Root repo README with the day-by-day table □ LinkedIn post or blog published □ Repo pushed to GitHub: ml-dl-fundamentals/
You started with arrays. You end with attention. Ship it.
⚡ Quick Recap — Days 14–15
10.1 One-Page Recap
| Topic | One-line memory |
|---|---|
| NumPy/Pandas | Vectorize; clean data first |
| Regression | Line fit; minimize MSE |
| Gradient Descent | Step downhill: w -= lr·grad |
| Overfitting | Memorizes train, fails on new data |
| Regularization | Penalize big weights (L1 sparse, L2 small) |
| Cross-Validation | Rotate validation folds for reliable scores |
| Metrics | Precision = right when yes; Recall = found all yes |
| Trees/Forest | Bagging cuts variance |
| Boosting | Sequentially fix errors |
| Neural Net | Layers of activation(XW+b) |
| Backprop | Chain rule backward |
| PyTorch | Automates the 5-step loop |
| Topic | One-line memory |
|---|---|
| CNN | Filters exploit image structure |
| Attention | softmax(QKᵀ/√d)V |
| Transformer | Attention + FFN + residual + LayerNorm, stacked |
10.2 Capstone Framework
Pick one model (recommended: MNIST CNN, or Attention from Day 13) and rebuild without tutorials. Then explain using this template:
- Problem: What are we solving? Why does it matter?
- Input / Output: shapes and meaning (e.g.
(B,1,28,28) → (B,10)) - How the model works: layers, data flow, diagram
- Loss function: which one and why
- Training: optimizer, lr, batch size, epochs, regularization
- Evaluation: metrics, curves, error analysis (confusion matrix)
Feynman test: Can you explain it to a friend with no ML background in 5 minutes?
10.3 Clean Code Checklist
□ Functions/classes, no copy-paste blocks □ Set seeds: torch.manual_seed(42), np.random.seed(42) □ Meaningful names, docstrings □ requirements.txt □ Notebook runs top-to-bottom □ No hardcoded local paths □ .gitignore (data, checkpoints, __pycache__)
10.4 README Template
# Project Title
One-sentence summary.
## Problem
## Dataset
## Approach
(model diagram / pipeline)
## Results
| Model | Metric | Score |
(screenshots: loss curves, confusion matrix)
## Key Learnings
## How to Run
pip install -r requirements.txt
python train.py
## Future Work10.5 LinkedIn Post / Blog Structure
- Hook: “I trained a CNN in PyTorch in 15 days. Here’s what surprised me.”
- What I built (1–2 lines + link)
- 3 key learnings (specific, technical)
- Result (number or image)
- What’s next (GenAI/RAG path)
- Hashtags:
#MachineLearning #DeepLearning #PyTorch #AI
Blog outline: Problem → Intuition → Math → Code → Results → Mistakes made.
10.6 Repo Structure
ml-dl-fundamentals/
├── Day-01-NumPy-Pandas/
├── ...
├── Day-10-Capstone/
└── README.md ← index with one-line summary + link per dayBeginner checkpoint — Days 14–15 ■ Explain the day’s main idea without using the textbook wording. ■ Write one tiny example from memory. ■ Identify one common failure mode. ■ Explain how today’s topic connects to the next day. ■ Write one question you still cannot answer.