What data scientists do: the workflow
See the six stages of a data science project, from framing a business question to deploying a result, using one worked example from a coffee shop. Learn where time really goes and the mistakes that sink beginners.
Data science is not just training models. Most of a data scientist's week goes into understanding a business question, finding and cleaning data, and explaining results. Model training is often the shortest step.
This lesson gives you the map of the whole workflow. Every later lesson in this track fits into one of its stages.
The workflow in six stages
The data science workflow
| Stage | Question it answers | Typical tools |
|---|---|---|
| 1. Frame the problem | What decision will this help, and how will we know it worked? | A conversation, a one-page brief |
| 2. Get the data | Where is the data and are we allowed to use it? | SQL, CSV files, APIs, Kaggle datasets |
| 3. Clean and explore | What is in the data and what is wrong with it? | pandas, matplotlib, seaborn |
| 4. Model | Which method can learn the pattern? | scikit-learn, PyTorch |
| 5. Evaluate | Is it good enough on data it has never seen? | Metrics, train/test split |
| 6. Communicate and deploy | Who will use this, and how? | Charts, a report, an API, a dashboard |
The stages are a loop, not a line. Exploring data often shows that the question was wrong. A weak model often sends you back to get better data. Expect to go round more than once.
A worked example: Brew & Bean
Brew & Bean, a coffee shop in Indiranagar, Bengaluru, is wasting milk and pastries on slow days and running out on busy ones. The owner asks: "Can you tell me how much to prepare tomorrow?"
Here is how each stage looks for this one question.
One question, six stages
- 1Frame it. The vague wish is "prepare the right amount". Make it measurable: predict the number of cups sold tomorrow. Success means the prediction is usually within 10 cups, and food waste falls over the next month.
- 2Get the data. Order history from the shop's billing system (one row per order), plus the weather forecast and a list of promotion days. Check that you are allowed to use it and that customer phone numbers are removed.
- 3Clean and explore. Load the data in pandas. Find missing temperatures, one day when a billing glitch recorded 900 cups, and a clear pattern: more sales on hot days and weekends.
- 4Model. Start with the simplest model: a linear regression using temperature, weekend and promotion. Compare it to a naive guess (always predict the average).
- 5Evaluate. Test on the last 30 days, which the model never saw. Report the average error in cups, not just a score.
- 6Communicate and deploy. Give the owner a simple table: expected cups for the next 7 days. Later, automate it as a small API or a dashboard.
Notice what the first stage did. "Prepare the right amount" became a number you can predict and measure. Many failed projects skip this step and build a model that nobody can use.
Types of problems
Common problem types
| Type | What you predict | Example |
|---|---|---|
| Regression | A number | Cups sold tomorrow, house price |
| Classification | A category | Will this customer churn? Is this email spam? |
| Clustering | Groups, with no labels | Segments of customers who buy alike |
| Forecasting | A number over future time | Next month's revenue |
| Recommendation | Which items a person may like | Suggest a croissant with a cappuccino |
Also ask whether you need machine learning at all. If a simple rule or a SQL query answers the question ("which item sells most on Mondays?"), use that. A model is worth its cost only when a pattern is too complex for a rule.
Roles around data
Roles
| Role | Main focus |
|---|---|
| Data analyst | Answers business questions with SQL, dashboards and charts |
| Data scientist | Analysis, statistics and models; often frames the problem too |
| ML engineer | Puts models into production and keeps them running |
| Data engineer | Builds the pipelines that deliver clean data (see the Data Engineering track) |
Titles overlap between companies. Read the job description, not just the title. This track prepares you for the analysis and modelling parts; deployment is covered in *MLOps basics: from notebook to deployed model* at the end of the track.
Common mistakes
Beginner traps
Three mistakes that waste weeks: (1) starting with a model before you understand the data, (2) testing the model on the same data it learned from, which gives a falsely high score, and (3) using information the model would not have at prediction time, called data leakage. Later lessons show how to avoid each.
Write the problem statement in one sentence before opening a notebook: "Predict [what] for [whom] using [data], so that [decision]." If you cannot write it, you are not ready to code.
I work as a junior data scientist. My manager said: "[paste the vague request]".
Help me turn this into a clear data science problem. Give me:
1. A one-sentence problem statement
2. Whether it is regression, classification, clustering or something simpler like a SQL report
3. The data I would need, with example columns
4. One metric to judge success, in plain business terms
5. Three risks (for example missing data or leakage)
Ask me up to 3 questions first if something important is unclear.What to expect: A good answer picks a measurable target, names concrete columns and warns about at least one data problem. Check that the metric matches the business decision.
How this comes up in interviews
Interviewers often ask "Walk me through a project" or "How would you predict X?". A strong answer follows the workflow: the business goal, the data and its problems, a simple baseline, how you evaluated, and what you would do next. Mention the baseline and the data cleaning; many candidates only talk about the algorithm.
Try it: Frame your own problem
Pick a place you know: your college canteen, a local kirana shop, or your hostel mess. Write the six stages for one question, such as "How many chapatis should the mess make on Friday?". For each stage write one or two lines: what data you would collect and how you would check success.
Key takeaways
- Data science is mostly framing, data cleaning and communication; modelling is one stage.
- The stages are: frame, get data, clean and explore, model, evaluate, communicate and deploy, and you will loop back.
- Turn vague wishes into a measurable prediction target before writing code.
- Always start with a simple baseline and test on data the model has not seen.
- Sometimes a SQL query or a simple rule is the right answer, not a model.
Check yourself
Finishing the quiz marks the lesson as done.
1. A shop owner says: "Help me sell more." What should you do first?
2. Which problem is classification?
3. Why do we test a model on data it has never seen?
4. Which statement best describes how the workflow is run in practice?
5. A simple SQL query already answers the owner's question. What is the best choice?
Saved on this device. Create a free account to keep it on every device.
Practise this with Sam
A real mock interview, out loud, with an honest scorecard. Free to start.