The data stack
0/3Foundations
Pandas is a spreadsheet you drive with code. SQL asks a database for exactly the rows you need. NumPy is the fast arrays underneath both.
Beginners often skip SQL and regret it later.
Data cleaning
0/1Bad data in, bad answers out, no matter how clever the model.
Real data has missing values, typos and wrong formats. Cleaning makes it trustworthy before you ask it anything.
Exploratory data analysis
0/1Before you model
Get to know a dataset before trusting it: count things, plot things, look for anything odd.
The cheapest step in the process, and it catches mistakes that would quietly ruin a model weeks later. Expect to loop between cleaning and EDA several times.
- Shape first
- Rows, columns, types, missing values. Look at raw rows too.
- Distributions
- Skew, double peaks, impossible values.
- Relationships
- Correlation shows things move together, not that one causes the other.
- Leakage
- A column that secretly contains the answer. Perfect in testing, useless in reality.
Machine learning
0/2The big one
A model is a maths function. Training searches for the version of that function that best matches the answers you already know.
Feed it inputs with known answers, let it adjust its formula until outputs match, then trust it on new inputs. Reinforcement learning, the third branch, waits until Phase 04.
- Weights
- How much each input matters.
- Bias
- A baseline that shifts the whole result up or down.
- Supervised
- Every training example comes with the right answer.
- Unsupervised
- Finds structure in data with no answers attached.
Data ethics
0/1A model trained on biased data repeats that bias at scale, with confidence.
Ask who could be harmed and how you'd find out before you ship anything. This is part of the job, not an extra.
Deployment & MLOps
0/2Shipping it
A model in a notebook is a sketch. MLOps is everything around it: build it reproducibly, serve it, and notice when it drifts.
Deployment turns a model into something other people can use: packaged, behind an address your app can call, and watched afterwards.
- Serving
- Wrap the model in a small API. FastAPI plus Docker is the standard first stack.
- Reproducibility
- Same data, code and settings, versioned, give the same model.
- Monitoring & drift
- The world changes and your model doesn't. Watch inputs and predictions.
- Pipelines
- Collect, clean, train, evaluate, deploy, on a schedule.
Embeddings & vector databases
0/1Newer piece
An embedding turns meaning into coordinates. A vector database finds the nearest points fast.
Similar things land close together, so “find what's nearest” powers semantic search, recommendations and the retrieval step in RAG.
- Similarity
- Cosine similarity is the usual measure of closeness.
- Chunking
- How you split documents is the biggest lever on retrieval quality.
- ANN indexes
- Approximate search gives up a little accuracy for huge speed.
Build this
Pick a Kaggle dataset you find fun. Clean it, explore it, train a scikit-learn model, then serve predictions from a FastAPI endpoint inside Docker.
Check yourself before moving on
Tick what you can really do, without looking it up. Anything you can't yet is where to go back.