How Does Data Science Work in Real Companies Today

Framing — the part tutorials skip:
"A manager says: 'We think churn is going up.
Can you look into it?'

The real work:
- What does 'churn' mean here?
- Which customers count?
- What time window?
- What decision will this change?
- What would success look like?

You leave with a measurable question:
'Among paying customers active in the
last 90 days, which segments have the
highest 30-day churn rate, and what
drives it?'

That question is the project."

Cleaning — where most time goes:
"The data you get is never clean. It has:
- Nulls in fields that should be required
- Duplicate rows from a broken sync
- Timestamps in three timezones
- Free-text fields with typos
- Deleted users still present in joins

What you actually do:
- Profile the data before trusting it
- Write a cleaning script you can rerun
- Document every assumption
- Flag what you excluded and why

What you say to the stakeholder:
'The churn number is X, but it excludes
customers with missing signup dates —
about 4% of the base. Here is the impact
of including them.'

That honesty is the job."

The Baseline — why it matters:
"Everyone wants the best model. But the
first thing you should ship is a baseline
— a simple rule or a logistic regression.

Why:
- It gives you a number to beat
- It is explainable to stakeholders
- It is easy to deploy
- It can already be useful

Example:
'Customers who have not logged in for 14
days churn at 3x the base rate.'

That rule might be 80% as good as a fancy
model — and it can ship in a week, not a
quarter.

The model can come later. The baseline
is what buys you time."

The Impact — why it matters:
"A model that nobody uses has no impact.
A dashboard that nobody reads has no
impact. The job is not to build — it is
to change a decision.

What to do:
- Ask who will act on this result
- Ask what they will do differently
- Ask what would make them trust it
- Involve them before you build

Example:
Instead of delivering a 0.91 AUC churn
model, you deliver:
'Here are the top 5% of customers most
likely to churn this month. Here is the
list. Here is what retention did when we
called them last time.'

That is impact. The model is secondary."

Weak Answer:
"I built a churn model. It got 0.91 AUC.
I used XGBoost with feature engineering."

Interviewer: "What did the business do
with it?"
You: "I'm not sure. I just handed it over."

Interviewer: "What was the messy part?"
You: "The data was pretty clean."

The conversation ends here.

Strong Answer:
"The question was: which paying customers
are likely to churn in the next 30 days,
and what can retention do about it?

The data was messy — 40% of signup dates
were null, and three systems disagreed on
'active'. I documented the assumptions and
excluded the 4% we couldn't classify.

I shipped a logistic regression baseline
first — it was 80% as good as XGBoost and
explainable. Then I upgraded to XGBoost
and compared.

The top 5% list went to the retention team,
who used it for calls. Churn in that group
dropped 12% over the quarter. The model
was secondary — the list was the product."

The conversation continues.