Data Analyst Intern

When people hear the term Data Scientist, they usually imagine machine learning models, recommendation systems, and endless discussions about hyperparameters.

My internship at Axtria taught me something completely different.

Sometimes, the most impactful data science project isn't about building a model.

Sometimes, it's about taking a process that consumes two days every week and teaching a computer to finish it before your second cup of tea.

Welcome to the world of pharmaceutical call planning.

A call plan determines:

• Which doctors should be visited.
• Which products should be promoted.
• How frequently doctors should be contacted.
• Which sales representative should visit them.
• When those visits should occur.

Imagine having:

• Millions of prescription records.
• Thousands of physicians.
• Hundreds of territories.
• Multiple pharmaceutical products.
• A mountain of business rules.

Now imagine updating all of that every single week.

That is a call plan.

In a nutshell: Pharmaceutical call planning combines physician targeting, territory alignment, sales force optimization, and business constraints to determine how field representatives should engage with healthcare professionals. Modern call planning systems rely heavily on data analytics because physician accessibility, treatment protocols, and business priorities change continuously.

The project focused on automating AstraZeneca's Q2 2024 call planning process.

The existing workflow was repeated every week.

New data arrived.

Business rules changed.

Physicians moved between segments.

Territory assignments evolved.

Then the entire process started again.

The pipeline processed more than 100 GB of data stored in SQL Server and transformed it into a configurable weekly call plan.

Working with datasets of that size teaches you a valuable lesson:

A bad query doesn't fail.

It simply decides to ruin your entire afternoon.

The workflow was implemented entirely in Python using Pandas, with configuration files controlling weekly execution parameters.

The weekly call plan wasn't generated from a single database.

That would have been far too convenient.

Instead, the pipeline had to combine data from three independent sources:

1. A centralized Big DataLake containing large-scale historical datasets.
2. A Microsoft SQL Server database containing operational data.
3. Multiple Excel files, sometimes several gigabytes in size, shared by different business teams.

Each source had its own schema, naming conventions, update schedules, and data quality challenges.

The first challenge wasn't analytics.

It was convincing all three sources to speak the same language.

Combining data from three different sources sounds straightforward.

It isn't.

A physician identifier stored as an integer in one system might appear as a string in another.

Column names that should match often don't.

Missing values suddenly become everyone's problem.

Excel files introduce their own surprises.

And spreadsheets measuring several gigabytes in size have a special ability to make laptops question their life choices.

Before any analysis could begin, the data had to be:

• Extracted.
• Validated.
• Standardized.
• Merged.
• Cleaned.

Only then could the business logic be applied.

The pipeline wasn't designed to run once.

It was designed to run every week.

Business users could modify configuration parameters without rewriting the codebase, allowing the same pipeline to adapt to changing business requirements.

The goal wasn't simply to automate a report.

The goal was to build a reusable system that could continuously generate call plans as new data became available.

This internship completely changed my understanding of data science.

I learned about:

• Pharmaceutical analytics.
• Sales force effectiveness.
• Physician segmentation.
• Large scale data processing.
• Workflow automation.
• Distributed computing.

Most importantly, I learned that not every data science problem needs a neural network.

Sometimes, the most valuable algorithm in the room is the one that quietly saves 44.5 hours every week.

And yes, watching a 48 hour job finish before lunch is deeply satisfying.