[reference] — from the roadmap.sh ai-data-scientist and data-analyst roadmaps. Orientation for the whole track.
What a data scientist does
A data scientist turns data into decisions. The job is less “build a model that ships” and more “answer a question well enough that someone acts on it”: Is this new feature actually working? Why did churn spike? Which customers should we target? Will this change help or hurt? The deliverable is usually an analysis, an experiment result, or a model used for insight — often a notebook and a presentation, not a production service.
DS vs ML Engineer vs Data Analyst
The three blur in practice, but the emphasis differs:
| Role | Optimizes for | Typical output | Depth of stats | Ships production software? |
|---|---|---|---|---|
| Data Analyst | describing what happened | dashboards, reports | descriptive | no |
| Data Scientist | explaining why & predicting | experiments, models, recommendations | deep (inferential + causal) | sometimes |
| ML Engineer | a reliable model in production | a deployed, monitored model | applied | yes |
The Data Scientist sits between the analyst (who reports) and the ML engineer (who ships), leaning hardest on statistics and experimental rigor — the skills that let you say “this result is real, not noise” and “this actually caused that.” Many real jobs are a blend; the split is about where the weight sits.
The analytics spectrum
A useful frame for what kind of question you’re answering — each level is harder and more valuable:
- Descriptive — what happened? Summaries, dashboards, descriptive stats. (“Sales fell 12% last quarter.”)
- Diagnostic — why did it happen? Drilling in, correlation, segmentation, EDA. (“Because churn rose in one region.”)
- Predictive — what will happen? Models forecasting the future (classical ML). (“These accounts are likely to churn.”)
- Prescriptive — what should we do? Recommending actions, optimization, experiments to test them. (“Offer this segment a discount — we A/B tested it.”)
Most business value is unlocked by moving up this ladder — and the higher rungs need the statistical and causal rigor this track is about.
The data science workflow
Real DS work is a loop, and (like ML engineering) it’s mostly not modeling:
frame the question → get & clean data → explore (EDA) → analyze / model
→ validate (is this real? is it causal?) → communicate → inform a decision
- Framing is underrated — a precisely stated question (“does X cause a lift in Y, for whom?”) determines everything downstream. A vague question yields a useless answer.
- Data cleaning & EDA are most of the time (working with data, EDA).
- Validation is what separates a data scientist from someone who found a pattern in noise — inferential statistics and causal reasoning.
Communication is half the job
The best analysis is worthless if no one acts on it. A data scientist must translate a statistical finding into a decision-maker’s language — a clear visualization, an honest statement of uncertainty (“we’re 95% confident the lift is 2–5%”), and a recommendation. Two failure modes to avoid: false certainty (presenting a noisy estimate as fact) and drowning the signal (a 40-slide deck when one chart would do). This “storytelling with data” skill is genuinely part of the role, not a soft add-on.
The toolkit
Python (pandas, the ML stack) and increasingly notebooks are standard, but data science leans on R more than the other paths — it’s built for statistics and has unmatched packages for inference, modeling, and visualization (ggplot2). SQL is essential (the data usually lives in a database), and analysts/DS often use BI tools (Tableau, Power BI) and even Excel for reporting (visualization).
Related
- Descriptive Statistics — the first analytical tool
- Inferential Statistics — the “is it real?” toolkit
- AI/ML course map — the three paths compared
Software engineering
Infrastructure & security
Computer science
Hardware & engineering