How to become an AI/ML engineer without a CS degree

You do not need a computer science degree to work as an AI or ML engineer. You need Python, enough statistics and linear algebra to understand what a model is doing, and three or four projects on GitHub that show you can take a dataset from raw rows to a working, measured, deployed model.

This guide is for graduates from any branch, including commerce, mechanical, electronics and science, and for working professionals in non-coding roles. It covers what the role involves, the maths that matters, a monthly project method, how interviews test you, and a four-month plan that mirrors our AI & ML Engineer track.

What the job actually involves

Most of an ML engineer's week is not modelling. A typical project runs through five stages, and the model is the shortest one.

• Data. Finding it, pulling it from a database or an API, and checking what is missing, wrong or duplicated. Expect half your time here.
• Features. Turning raw columns into inputs a model can use. A timestamp becomes day-of-week and days-since-last-order. Text becomes word counts or embeddings.
• Models. Picking a model family, training it and tuning a few settings. With scikit-learn this is often ten lines of code.
• Evaluation. Splitting data properly, choosing a metric that matches the business question, and comparing against a simple baseline. This is where most juniors fail interviews.
• Deployment. Wrapping the model in an API with FastAPI, putting it in a container, and checking whether it still works next month when the data has drifted.

Google's "Rules of Machine Learning" guide says most of the problems you will face are engineering problems. That is good news if you do not have a CS degree. Engineering habits are learned by doing, and nobody checks your transcript when your API returns the right answer.

The maths you need, and the maths you can skip for now

You need enough maths to read a model's output and know when it is misleading you.

• Statistics. Mean, median, variance, what a distribution looks like, and why 95% accuracy means nothing on a dataset where 95% of rows share one label.
• Probability. Conditional probability and Bayes' rule, because a classifier is a probability estimator with a threshold on top.
• Linear algebra. Vectors, matrices, dot products and matrix multiplication. A neural network layer is a matrix multiplication followed by a non-linear function. Once you see that, deep learning stops being magic.
• Calculus. What a gradient is and why following it downhill finds a minimum. You will not differentiate by hand at work. You need the picture.

Skip for now: proofs, measure theory, convex optimisation and deriving backpropagation on paper. Those matter for research roles and for a master's degree. They do not matter for shipping a churn model. In our Foundations course the maths is one module, taught as Python you run rather than symbols you memorise.

Python first, and only Python

Pick one language and stay with it. NumPy, Pandas, scikit-learn, PyTorch, TensorFlow, Hugging Face and LangChain are all Python-first, and AI/ML job posts list Python before anything else. Learning R or Julia first is a detour.

Learn it in this order: variables, lists, dictionaries and functions; then Pandas (read a CSV, filter, group, join); then plotting with matplotlib; then Jupyter notebooks as your lab book. Leave classes, decorators and type hints until a project forces you to need them. An hour a day for four weeks is enough to start the first project, and the first project teaches you the rest.

One project a month

The method we teach is one project a month, scoped on day one.

1. Week 1. Pick a public dataset and write the question in one sentence. "Will this customer buy?" is a question. "Customer analytics" is not.
2. Week 2. Clean the data and build the simplest baseline that could work: predict the majority class, or last month's value. Record its score.
3. Week 3. Beat the baseline with one or two model families. Tune a little. Write down every decision in the README.
4. Week 4. Deploy it as a small app with Streamlit or FastAPI, record a two-minute screen capture, and push to GitHub with a README a recruiter can read in 60 seconds.

Four months of this gives you four finished projects. Finished matters more than impressive. A deployed logistic regression with a measured lift over the baseline beats an unfinished transformer.

Portfolio over certificates

A certificate proves you watched. A repository proves you built. Hiring managers open GitHub links and rarely open certificate PDFs. Put certificates on one line of the resume and spend the space on three projects, each with the question, the dataset, the metric with your number against the baseline, and a link to the live demo.

A degree in another subject is an asset if you use it. A mechanical engineer who models sensor failures, a commerce graduate who models loan defaults, a pharmacy graduate who classifies adverse-event reports: each has a story in interviews that CS graduates do not. Make at least one of your four projects about your own field.

How interviews test you

Rounds vary by company, but a first AI/ML role in India usually has four parts.

1. Python and SQL screen. Write a function, a Pandas transform, a join with a group by. Often online and timed.
2. ML fundamentals. Bias and variance, overfitting, cross-validation, precision against recall, why accuracy misleads, how gradient boosting differs from a random forest, what regularisation does. Expect to draw a confusion matrix.
3. Your project. Twenty to thirty minutes on one repository. Why this dataset, how you split it, what leaked, what you would do with ten times the data. This is where the DECISIONS file pays off.
4. A case. "We want to predict which orders will be returned. How would you start?" They want data sources, a baseline, a metric and the risks, in that order.

Nobody asks about your degree in rounds two to four. Round one filters on Python, which is why Python comes first.

A 4-month plan from zero

This mirrors the two courses in our AI & ML Engineer track: Foundations in months one and two, Applied AI & Deployment in months three and four. Each course is 48 live hours over two months, so budget about six hours a week of class and the same again for your project.

If you are self-studying, the same order works. The datasets and definitions of done for each project are in 7 machine learning projects that get you interviews.

When a bootcamp is worth it

Self-study works for people who can hold a schedule alone for four months and get unstuck without help. Most people cannot, and that is the honest reason live courses exist. A course is worth paying for when it gives you four things that are hard to get alone: a fixed timetable, someone who reviews your code and your README, mock interviews with people who hire, and referrals into open roles. It is not worth paying for a library of recorded videos. YouTube already has those.

Before you pay anyone, ask who teaches and what they build at work, how many projects you will finish, whether you present your capstone to anyone, and what the placement assistance includes in writing.

What to do next

• Install Python and Jupyter this week and load one CSV into Pandas.
• Pick your first dataset from the seven projects article and write the question in one sentence.
• If you want the structured route, read the AI & ML Engineer track page, or talk to a counsellor.
• We also run a free AI & ML masterclass, where you train a first model live and see data, features, model, evaluation and app in three hours.

Sources

• Google, Rules of Machine Learning
• scikit-learn user guide: cross-validation and evaluating estimator performance
• PyTorch tutorial: transfer learning for computer vision
• FastAPI documentation