Data and artificial intelligence · 5 stages · about 54 weeks
Data scientist / machine learning engineer roadmap
0/25 tasks done (ticks are kept on this device)
- 1
Beginner · about 8 weeks
Python and programming basics
What to learn
Python syntax: Python is the main language of data science.
Learn variables, numbers, strings, if/else, for and while loops, and functions. Write small programs every day, like a marks-to-grade converter or a GST calculator. Clear, simple code is more important than clever code.
Lists, dictionaries, sets and tuples: These hold the data your programs work on.
A list is an ordered collection, a dictionary stores values by key, a set holds unique items and a tuple is a list that cannot change. Learn list comprehensions, a short way to build lists. You will use dictionaries all the time when working with JSON data.
Files and modules: Real data comes from files and real code uses libraries.
Learn to read and write text and CSV files, and to import modules like math, random and datetime. Install packages with pip, and use a virtual environment so each project has its own packages. This keeps your projects from breaking each other.
Functions and clean code: Good functions make analysis easy to repeat and test.
Each function should do one job and have a clear name. Use docstrings to say what it does, and avoid copying the same code in many places. Learn basic error handling with try and except.
Classes basics: ML libraries are built with classes, like model.fit().
A class bundles data and functions together, like a BankAccount with a balance and a deposit method. You do not need deep object-oriented design, but you must read class-based code comfortably. scikit-learn and PyTorch both use classes heavily.
Jupyter and Colab: Most data work is done in notebooks.
A notebook mixes code, results, charts and notes in one page. Google Colab runs notebooks free in the browser and even gives you a GPU for small deep learning tasks. Keep notebooks tidy and run them top to bottom before sharing.
Git and GitHub: Your GitHub is your data science portfolio.
Git saves versions of your code and notebooks. GitHub lets employers see your projects. Never upload passwords, API keys or private data.
Tasks
Project
A command-line tool that reads a CSV of daily shop sales, calculates daily, weekly and monthly totals, finds the best-selling items, and saves a summary report to a new file.
📘 Learn it here, free: Python Programming course
Free resources
- 2
Beginner · about 12 weeks
Maths for ML and working with data: NumPy, pandas, SQL and charts
What to learn
Linear algebra: Data and models are stored and computed as vectors and matrices.
A vector is a list of numbers, like one customer's age, income and city code. A matrix is a table of numbers, like many customers. Learn matrix multiplication, the dot product and the idea of dimensions; you need intuition more than proofs.
Probability: ML is about making good guesses when things are uncertain.
Learn probability, conditional probability and Bayes' rule with simple examples, like the chance a message is spam given it contains the word "offer". Learn common distributions such as normal and binomial. Many model outputs are probabilities, so you must read them correctly.
Statistics: Statistics tells you if a result is real or just luck.
Learn mean, median, variance, standard deviation, sampling, confidence intervals and hypothesis tests at a practical level. Understand that correlation does not prove causation. These ideas come up in every interview and every real project.
Calculus intuition: Models learn by slowly reducing their errors.
A derivative tells you how much an output changes when an input changes a little. Gradient descent uses this to adjust model weights step by step to reduce error. You rarely calculate by hand, but you must understand the idea.
NumPy: NumPy does fast maths on large arrays of numbers.
NumPy arrays are much faster than Python lists for maths. Learn array creation, shape, indexing, slicing and broadcasting, which is how NumPy combines arrays of different shapes. Almost every ML library uses NumPy underneath.
pandas: pandas is the main tool for cleaning and exploring tables.
A DataFrame is like an Excel sheet in Python. Learn to load files, select and filter, handle missing values, groupby, merge and work with dates. Most real projects spend more time here than on modelling.
SQL: Company data lives in databases.
Learn SELECT, WHERE, GROUP BY, JOIN and window functions. Many data science interviews include a SQL round. Pull only the data you need from the database, then continue in pandas.
Visualisation: Charts help you find patterns and explain results.
Use Matplotlib and seaborn for histograms, scatter plots, box plots and line charts. Use charts to explore data before modelling, and to explain results after. Every chart needs a title, labels and one clear message.
Tasks
Project
An exploratory data analysis of a real dataset (for example Indian used-car prices or house rents): cleaning, statistics, 8 clear charts and a short write-up of what drives the price, published on Kaggle or GitHub.
Free resources
- 3
Intermediate · about 12 weeks
Machine learning with scikit-learn
What to learn
What machine learning is: You need the big picture before the details.
In supervised learning the model learns from examples with answers, like past loans marked repaid or not. In unsupervised learning it finds groups or patterns without answers. Regression predicts a number, like a price; classification predicts a category, like spam or not spam.
Train, validation and test splits: A model must be judged on data it has never seen.
Split data into a training set to learn from, a validation set to tune choices and a test set used only once at the end. Cross-validation repeats the split several times for a fairer score. Testing on training data gives falsely good results.
Core algorithms: A few algorithms solve most real business problems.
Learn linear and logistic regression, decision trees, random forests, gradient boosting (XGBoost or LightGBM), k-nearest neighbours and k-means clustering. Understand roughly how each works and when to use it. On table data, gradient boosting is often the strongest choice.
Model evaluation metrics: The wrong metric can make a bad model look good.
For regression, use MAE and RMSE. For classification, learn accuracy, precision, recall, F1 and ROC-AUC, and the confusion matrix. If only 1 in 100 transactions is fraud, a model that always says "not fraud" is 99 percent accurate and still useless.
Overfitting and underfitting: This is the most common reason models fail in real life.
An overfit model memorises the training data and fails on new data; an underfit model is too simple to learn the pattern. Regularisation, more data, simpler models and early stopping help. Compare training and validation scores to spot the problem.
Feature engineering: Good features often matter more than the choice of model.
Features are the inputs a model learns from. Create useful ones, like day of week from a date or price per square foot. Encode categories, scale numbers when needed, and handle missing values in a consistent way.
Pipelines and data leakage: Leakage gives great test scores and a useless real model.
Leakage happens when information from the test data or the future sneaks into training. scikit-learn Pipelines keep cleaning and modelling steps together so they are applied correctly. Always ask: would I really have this information at prediction time?
Hyperparameter tuning and explainability: Tuning improves models and explanations build trust.
Hyperparameters are settings you choose, like tree depth. GridSearchCV and RandomizedSearchCV try many settings for you. Feature importance and SHAP values help you explain which inputs drive a prediction.
Tasks
Project
A loan default or customer churn prediction project: clear problem statement, EDA, feature engineering, a scikit-learn pipeline, at least 3 models compared with the right metrics, an explanation of the main drivers, and an honest section on limits and bias.
Free resources
- 4
Expert · about 12 weeks
Deep learning with PyTorch
What to learn
Neural network basics: Deep learning powers image, speech and language models.
A neural network is layers of simple units that each multiply inputs by weights and pass the result on. Training means adjusting the weights with backpropagation and gradient descent to reduce a loss. Learn activations, loss functions, optimisers and learning rate.
PyTorch tensors and autograd: PyTorch is the most used deep learning library in research and industry.
A tensor is like a NumPy array that can run on a GPU. Autograd works out gradients for you automatically. Learn to move data and models to the GPU in Colab or Kaggle notebooks.
The training loop: Every PyTorch project uses the same basic loop.
Load data in batches with a DataLoader, run the model, compute the loss, call backward and step the optimiser. Track training and validation loss every epoch. Save the best model checkpoint, not just the last one.
Computer vision with CNNs: Image tasks are common: quality checks, documents, medical scans.
Convolutional neural networks (CNNs) find patterns like edges and shapes in images. Use data augmentation, like flips and crops, to get more from less data. Transfer learning, starting from a model trained on millions of images, usually beats training from zero.
NLP and transformers: Transformers power modern language models.
Transformers use attention to understand how words in a sentence relate. Hugging Face gives you pre-trained models you can fine-tune for tasks like sentiment or classifying support tickets, including in Hindi. Learn tokenisation, embeddings and fine-tuning at a practical level.
Working with large language models: Many ML jobs now include LLM work.
Learn prompting, embeddings and retrieval-augmented generation (RAG), which means giving the model the right documents before it answers. Learn to evaluate outputs with test sets, not just by looking at a few examples. Know when a simple scikit-learn model is cheaper and better.
Experiment discipline: Deep learning has many settings, and it is easy to fool yourself.
Change one thing at a time and record every run. Fix random seeds so results can be repeated. Always compare against a simple baseline model.
Tasks
Project
A deep learning project end to end: collect or choose a dataset, fine-tune a pre-trained vision or language model in PyTorch, compare it with a simple baseline, explain the errors it makes, and share the notebook and results on GitHub and Kaggle.
Free resources
- 5
Expert · about 10 weeks
MLOps, deployment and building a career
What to learn
Experiment tracking with MLflow: Teams must know which model, data and settings gave which result.
MLflow records parameters, metrics and model files for every run. Its model registry stores versions and marks which one is in production. This makes your work repeatable and easy to review.
Serving models with an API: A model is only useful when an app can call it.
Wrap your model in a small web API with FastAPI, so an app sends input and gets a prediction back. Check the input data and return clear errors. Measure how long each prediction takes.
Docker and cloud deployment: Docker makes your model run the same everywhere.
Package the API, the model and its libraries in a Docker image. Deploy it to a cloud service such as AWS, Google Cloud or Azure, or a simple platform for demos. Learn basic costs, because GPUs and large models can be expensive.
Monitoring and data drift: Models get worse over time as the world changes.
Data drift means new data looks different from training data, like buying patterns changing during festivals. Track input data, prediction distributions and real accuracy when answers arrive later. Plan when and how to retrain.
Data and ML pipelines: Real systems retrain and predict on a schedule.
A pipeline runs steps automatically: get data, clean it, train, evaluate and deploy if the new model is better. Tools such as Airflow or Prefect schedule these steps. Add tests for data quality, not only for code.
Responsible AI: Models can be unfair or leak private data.
Check whether your model works equally well for different groups, like regions, languages or genders. Protect personal data and follow India's Digital Personal Data Protection Act and company rules. Be clear about what your model cannot do.
Kaggle and portfolio: A strong public portfolio is how many people get their first data role.
Kaggle competitions and notebooks show that you can work on real problems. Build 3 to 4 deep projects instead of many copied tutorials, each with a clear README. Expect to start as an analyst or junior and keep building; that is a normal, healthy path.
Tasks
Project
An end-to-end ML product: a crop-price or demand forecast model (or another real problem) trained with a tracked pipeline in MLflow, served by a FastAPI app in Docker, with a simple web page to try it, a monitoring plan and a README that explains the business value and the limits.
Free resources