In data science, nothing slows progress like uncertainty — especially when you are not sure if a recent change in code or data has affected your results in unintended ways.
At Combine, we believe that reliable results are just as important as fast iteration. That is why we have refined how we use Continuous Integration (CI), building a workflow designed for the specific challenges of data science and AI development.

From Software Testing to Data Confidence

Traditional Continuous Integration focuses on testing and validating code. In data science, that is only part of the picture.
Data evolves, analysis pipelines change, and even small adjustments can alter results.
To handle that complexity, we use CI both as a way to verify code and to maintain data integrity.

In our setup, each project saves key outputs such as results, metrics, and statistics in structured CSV files.
On every merge request, these outputs are automatically compared to previous runs.
If a change in results appears, we can see it immediately — allowing us to determine whether the difference is expected or if it signals that something went wrong.

Shaping a Better Workflow

In a recent internal innovation project, our R&D team set out to make this process faster, clearer, and more consistent across projects.
We wanted a CI pipeline that engineers actually enjoy using — one that gives quick feedback when it matters and performs deeper validation when it is truly needed.

Here’s what we did:

Standardized analysis outputs

We introduced a shared format for how analyses are logged and stored, and documented how each analysis should be included in CI.
This consistency makes results easier to interpret and reuse between projects.

Made result changes transparent

A new visualization method now highlights changes in a clear and intuitive way, even when working with large and complex datasets.
Seeing what changed has never been easier.

Balanced speed with reliability

We defined two main types of checks:

• Data checks detect when datasets or inputs have changed. These run automatically and provide instant feedback.
• Code checks verify when analysis scripts produce different outputs. These are triggered before a merge, ensuring thorough validation without slowing everyday development.

The Results

The outcome is a Continuous Integration setup that fits how data scientists actually work. Pipelines run faster. Result changes are easier to understand.
Rules about what runs and when are now clear and predictable.

Beyond the technical improvements, this project also clarified our philosophy: Continuous Integration in data science is not just about testing code — it is about maintaining trust in our results as projects evolve.

At Combine, continuous improvement is part of how we operate. Our new workflow strengthens our ability to deliver reliable, reproducible, and high-quality results across all our data science and AI projects.

What This Means for Our Customers

For our customers, this improvement translates directly into higher confidence and faster delivery.
They can rely on our analyses to remain consistent over time, even as models, data, and systems evolve.
The new workflow allows us to detect unintended changes early, reducing risk and making every iteration more predictable.

It also helps us deliver results faster. With a clearer structure and faster feedback, we spend less time verifying and more time improving.
In short, our customers receive insights that are both trustworthy and delivered with greater efficiency — a combination that supports better decisions and more reliable AI systems in the real world.