Learning log

What I am learning, and how far along it is.

A 30-week path from Python to shipped AI systems, about fifteen hours a week. Every topic and every weekly project is ticked off as it is finished. Nothing here is a claim — it is a record.

Of the path

0 / 201

Steps finished

0 h

Of 462 h

Day streak

Programming foundations0 / 26

Data and math for ML0 / 12

Classical machine learning0 / 26

Deep learning0 / 19

Transformers and LLMs0 / 13

Building LLM apps0 / 14

Retrieval-augmented generation0 / 14

Agents and MCP0 / 14

Fine-tuning and evals0 / 14

Backend, Docker, CI/CD and deploy0 / 21

MLOps and system design0 / 21

Get hired0 / 7

Steps finished over time

Last twenty weeks · 0 steps in the last 7 days

Programming foundations

Python, Git, SQL and APIs. Every AI system is software first. Four weeks of Python, Git, SQL and APIs is the floor everything else stands on.

0 / 26 · 0%

Week 1 · Python basics

Install Python and VS Code, or open Google Colab2 h
Variables, numbers, strings, input and output2 h
if / else, for and while loops2 h
Lists, tuples, dictionaries and sets2 h
Functions, arguments and return values2 h
Read error messages instead of fearing them2 h
Ship: A command-line quiz or unit converter, pushed to GitHub with a README.4 h

Week 2 · Python for real work

Files, JSON and CSV2 h
try / except and your own error types2 h
Modules, packages, pip and uv environments2 h
Classes, objects and dataclasses2 h
Comprehensions, f-strings and type hints2 h
Small tests with pytest2 h
Ship: An expense tracker CLI that saves to JSON, with tests.4 h

Week 3 · Git, terminal and SQL

Terminal: navigate, pipes, environment variables2 h
Git: commit, branch, merge, fix conflicts2 h
GitHub: pull requests, issues, .gitignore2 h
SQL: SELECT, WHERE, GROUP BY, JOIN2 h
SQLite from Python, Postgres basics2 h
Ship: A SQL analysis of a real dataset, merged through a pull request.4 h

Week 4 · APIs and a tiny backend

HTTP methods, status codes and headers2 h
Call REST APIs with requests or httpx2 h
Keep API keys in a .env file, never in code2 h
async / await basics2 h
FastAPI routes and Pydantic models2 h
Ship: A FastAPI service that wraps a public API, with two endpoints and tests.4 h

Data and math for ML

NumPy, pandas, gradients, probability. Models eat data and speak in vectors. Two weeks to get fluent with NumPy and pandas and build intuition for the few math ideas ML uses.

0 / 12 · 0%

Week 5 · NumPy, pandas and plots

NumPy arrays, shapes, broadcasting2 h
pandas: load, filter, group, merge, pivot2 h
Missing values, dtypes, duplicates2 h
Charts that answer one question each2 h
Notebook workflow in Jupyter or Colab2 h
Ship: An EDA notebook on a real dataset: 5 charts, 5 written findings.4 h

Week 6 · The math you actually need

Vectors, dot products, matrix multiplication2 h
Derivatives, the chain rule, gradients2 h
Gradient descent, by hand and in code2 h
Probability, conditional probability, Bayes2 h
Distributions, mean, variance, sampling2 h
Ship: Gradient descent from scratch in NumPy fitting a line, with a loss plot.4 h

Classical machine learning

Algorithms, metrics, an end-to-end project. Before deep learning, learn how any model learns, overfits and gets measured. Interviews assume this vocabulary, and most real ML jobs still run on tabular data.

0 / 26 · 0%

Week 7 · How machines learn

Supervised vs unsupervised learning2 h
Train, validation and test splits2 h
Linear and logistic regression2 h
Loss functions and overfitting2 h
Accuracy, precision, recall, F1, ROC-AUC2 h
scikit-learn: fit, predict, score2 h
Ship: A price or marks predictor with a clean split and a metrics table.4 h

Week 8 · Trees, ensembles, clustering

Decision trees and random forests2 h
Gradient boosting: XGBoost, LightGBM2 h
Cross-validation and hyperparameter search2 h
k-means clustering and PCA2 h
Feature engineering and scaling2 h
Ship: A Kaggle submission with a notebook that explains every step.4 h

Week 9 · ML algorithms from scratch

kNN, naive Bayes and SVM: how each decides2 h
Linear and logistic regression with gradient descent in NumPy2 h
Decision tree split search (Gini, entropy) coded by hand2 h
k-means and DBSCAN: when each fails2 h
Bias-variance, regularisation (L1/L2), calibration2 h
Learning curves and what they tell you2 h
Ship: A small NumPy ML library (kNN, linear and logistic regression, k-means, decision tree) with unit tests and a comparison against scikit-learn.4 h

Week 10 · An end-to-end ML project

scikit-learn Pipeline and ColumnTransformer2 h
Imbalanced classes and the right metric2 h
Compare models, then study the errors2 h
Save and load models with joblib2 h
A small Streamlit or Gradio front end2 h
Ship: A churn or fraud predictor with a public demo link.4 h

Deep learning

From scratch, then PyTorch. Build neural networks from scratch, then in PyTorch, until you read training curves the way a doctor reads a scan.

0 / 19 · 0%

Week 11 · Neural nets from scratch

Neurons, layers, activation functions2 h
Forward pass and the loss2 h
Backpropagation is the chain rule2 h
Build a tiny autograd engine2 h
Learning rate, batches, epochs2 h
Ship: Your own micrograd-style engine training a small MLP.4 h

Week 12 · PyTorch fundamentals

Tensors, shapes and devices2 h
autograd and optimisers2 h
nn.Module, Dataset, DataLoader2 h
Training and validation loops, checkpoints2 h
Dropout, weight decay, early stopping2 h
Free GPUs on Colab and Kaggle2 h
Ship: A Fashion-MNIST classifier with training curves and 90%+ test accuracy.4 h

Week 13 · CNNs, transfer learning, embeddings

Convolutions and pooling2 h
Transfer learning from pretrained models2 h
Data augmentation2 h
Embeddings: things as vectors2 h
Why sequence models led to attention2 h
Ship: A classifier for your own photos, published as a free public Gradio demo.4 h

Transformers and LLMs

Attention, a tiny GPT, how LLMs are built. Open the black box. Build attention and a tiny GPT, then learn how real LLMs are pretrained, tuned and served.

0 / 13 · 0%

Week 14 · Attention and transformers

Tokenization and byte-pair encoding2 h
Token and position embeddings2 h
Self-attention and multi-head attention2 h
The transformer block: residuals, layer norm2 h
Decoder-only (GPT) vs encoder (BERT)2 h
Ship: A small character-level GPT trained on text you choose, with samples.4 h

Week 15 · How LLMs are built and run

Pretraining on web-scale text2 h
Instruction tuning (SFT) and RLHF / preference tuning2 h
Reasoning models and test-time compute2 h
Sampling: temperature and top-p2 h
KV cache, context windows, latency2 h
Open-weight models with Hugging Face and Ollama2 h
Ship: A comparison of 3 open models run locally on 20 of your own prompts.4 h

Building LLM apps

APIs, prompts, JSON, tool calling. Now you build. Call models through APIs, write prompts that hold up, force output into validated JSON and let models call your functions.

0 / 14 · 0%

Week 16 · LLM APIs and prompting

Chat APIs: system, user, assistant messages2 h
Prompts: clear task, examples, delimiters2 h
Stream responses into a UI2 h
Tokens, context windows and pricing2 h
Prompt caching and picking a model for cost2 h
Local models behind an OpenAI-compatible API2 h
Ship: A streaming chatbot that shows token count and cost for every reply.4 h

Week 17 · Structured output and tool calling

JSON mode vs schema-enforced output2 h
Pydantic models to validate everything2 h
Function / tool calling end to end2 h
Retries, timeouts and fallbacks2 h
Images and PDFs as inputs2 h
A fixed test set for your prompts2 h
Ship: A resume or invoice to JSON extractor, validated and tested on 20 files.4 h

Retrieval-augmented generation

Embeddings, vector search, RAG evals. Models do not know your documents. RAG finds the right passages and hands them to the model so answers are grounded and citable.

0 / 14 · 0%

Week 18 · Embeddings and vector search

What embeddings capture2 h
Cosine similarity and nearest neighbours2 h
Chunking: size, overlap, structure2 h
Vector stores: Chroma, pgvector, Qdrant2 h
Retrieve, augment, generate2 h
Cite a source in every answer2 h
Ship: A "chat with your notes" app that answers with citations.4 h

Week 19 · Better retrieval and RAG evals

Hybrid search: BM25 plus vectors2 h
Rerankers and top-k tuning2 h
Query rewriting and metadata filters2 h
Contextual retrieval2 h
Parsing PDFs and tables2 h
Retrieval hit-rate and faithfulness scores2 h
Ship: Hybrid search and reranking added, with before / after scores on 30 questions.4 h

Agents and MCP

The agent loop, LangGraph, MCP servers. Agents use tools in a loop until a task is done. Build them, connect them to real systems through MCP and keep them on a leash.

0 / 14 · 0%

Week 20 · Agent fundamentals

Workflow or agent: when autonomy helps2 h
The loop: model, tool, observation2 h
ReAct and planning2 h
Tool design: names, schemas, errors2 h
Short- and long-term memory2 h
LangGraph: state, nodes, edges2 h
Ship: A research agent with search and calculator tools, a step limit and full traces.4 h

Week 21 · MCP, multi-agent and safety

Model Context Protocol: tools, resources, prompts2 h
Build an MCP server and connect a client2 h
Human approval before risky actions2 h
Orchestrator-worker and other multi-agent patterns2 h
Permissions, step limits and cost limits2 h
Evaluate agents on real tasks2 h
Ship: An MCP server for a real API, used by an agent that asks before it acts.4 h

Fine-tuning and evals

LoRA, error analysis, tracing, security. Sometimes prompting and RAG are not enough. Learn when fine-tuning pays, then measure quality, trace every call and defend against prompt injection.

0 / 14 · 0%

Week 22 · Fine-tuning, when it pays

Prompting vs RAG vs fine-tuning: choose by problem2 h
Build a clean instruction dataset2 h
Fine-tune on a free Colab GPU with Unsloth or TRL2 h
Preference tuning (DPO) at a glance2 h
LoRA and QLoRA: train a small adapter2 h
Evaluate before and after, then publish to the Hub2 h
Ship: A LoRA fine-tune of a small open model for one narrow task, with a before / after eval.4 h

Week 23 · Evals, tracing and security

Error analysis: read real outputs, label failures2 h
Turn those failures into eval sets2 h
LLM-as-judge, and checking the judge2 h
Tracing with Langfuse or Phoenix2 h
Prompt injection and the OWASP LLM Top 102 h
Input and output guardrails2 h
Ship: An eval harness and tracing for your RAG app, run on every change.4 h

Backend, Docker, CI/CD and deploy

Django, containers, pipelines, production. A demo is not a product. Add a full web framework, package it, automate testing and deployment, and ship it where strangers can use it.

0 / 21 · 0%

Week 24 · Django and Django REST Framework

Django project layout, MTV, settings and apps2 h
Models, migrations and the ORM (select_related, prefetch_related)2 h
Admin, forms and authentication2 h
Django REST Framework: serializers, viewsets, permissions, pagination2 h
Calling your FastAPI model service from Django2 h
Django or FastAPI: when to pick each2 h
Ship: A Django + DRF app with login that stores user feedback and calls your FastAPI model service, with tests.4 h

Week 25 · Docker, CI/CD and cloud deploy

Dockerfile: multi-stage builds, layer caching, non-root user2 h
docker compose: app, Postgres, Redis, vector store, healthchecks2 h
GitHub Actions: lint, test, build and push an image on every merge2 h
Secrets, branch protection, blue/green and canary releases2 h
Deploy to Cloud Run, ECS or Render; environment config2 h
Kubernetes concepts: pod, deployment, service (read level)2 h
Ship: A GitHub Actions pipeline that tests, builds a Docker image and deploys your app to a public URL on every push to main.4 h

Week 26 · Ship to production

Serve your app with FastAPI and streaming2 h
Caching, rate limits, retries, timeouts2 h
Log cost and latency for every request2 h
Serving open models with vLLM or Ollama2 h
Load testing with Locust or k62 h
Health checks, logs and alerts2 h
Ship: A Dockerised app at a public URL with a cost and latency log and a load-test result.4 h

MLOps and system design

Tracking, serving, drift, design interviews. Models drift and teams need to track experiments. Design rounds appear in most interview loops, and ML design adds data, training, serving and evaluation.

0 / 21 · 0%

Week 27 · MLOps and model serving

Experiment tracking and a model registry (MLflow or Weights & Biases)2 h
Data and model versioning with DVC2 h
Pipelines: Airflow or Prefect; batch vs online inference2 h
Serving: ONNX, Triton, vLLM; batching and quantisation2 h
Monitoring: data drift, concept drift, alerts2 h
Shadow deploys, A/B tests and canary models2 h
Ship: MLflow tracking and a registry added to your ML project, the best model served behind FastAPI, and a drift check on a schedule.4 h

Week 28 · System design fundamentals

A framework: requirements, scale estimates, API, data model, design, trade-offs2 h
Load balancers, CDN, caches and their failure modes2 h
Databases: SQL vs NoSQL, replication, sharding, CAP2 h
Queues and streams (Kafka, SQS), idempotent consumers2 h
Rate limiter, consistent hashing, URL shortener2 h
Latency budgets and back-of-envelope maths2 h
Ship: Two written design docs with diagrams: a URL shortener and a rate limiter.4 h

Week 29 · ML and LLM system design

ML design steps: data, labels, features, training, serving, monitoring2 h
RAG chatbot for millions of users2 h
Recommendation feed: candidate generation then ranking2 h
Search ranking and fraud detection designs2 h
LLM gateway: routing, caching, fallbacks, cost control2 h
Failure modes: cold start, feedback loops, drift, hallucination2 h
Ship: Design docs for a RAG chatbot and a recommender, each practised aloud in a 35-minute mock.4 h

Get hired

Capstone, resume, mocks, applications. Turn the commit history into proof. Finish the capstone, show your work in public, rehearse the interview loop and apply every week.

0 / 7 · 0%

Week 30 · Portfolio, interviews and job hunt

Finish one capstone that solves a real problem2 h
README: problem, architecture, evals, demo GIF2 h
Resume: projects first, numbers in every line2 h
Post what you built on LinkedIn and X2 h
Record a two-minute demo video2 h
Mock interviews (coding, ML and system design), then apply every week2 h
Ship: Capstone live, 3 posts published, 30 applications sent, 3 mock interviews done.4 h