The 4 steps NEPA Framework to debug broken ML models
(without burning out)
Debugging machine learning models can feel like trying to untangle a mess of Christmas lights... blindfolded... with your boss pacing behind you.
It’s stressful.
It’s time-consuming.
And if you’re not careful, it will eat your entire weekend.
But there’s a better way.
I want to share a framework I stole from an unexpected place—elite fitness coaching—that has helped me:
Cut model debugging time in half
Improve my team’s confidence in our process
Build cleaner, more resilient ML pipelines
It’s called NEPA. And it might just save your sanity.
Let’s walk through each of the 4 steps:
1. Noticing: Slow down before diving in.
The biggest mistake I made as a data scientist? Jumping straight into logs.
Instead, take a breath. Zoom out.
Ask:
What recently changed in the data?
Did someone tweak the pipeline?
Has the environment (hardware, dependencies, external services) shifted?
Before touching a single line of code, investigate the context around the failure. This isn’t just technical sleuthing—it’s mindfulness applied to debugging.
2. Explaining: Form hypotheses out loud.
Once you’ve surveyed the system, don’t go solo.
Start articulating hypotheses—even if they sound dumb.
Grab a teammate and talk it through:
“What if it’s a timezone mismatch?”
“Maybe the retraining job silently failed?”
“Could this be a schema drift we missed?”
Talking builds clarity. It helps you move from reacting to reasoning. And it avoids the “silent rabbit hole” problem where you spend hours chasing the wrong thread.
3. Prescribing: Design a surgical fix.
Now you’re ready to act—but deliberately.
What does the fix actually require?
Retraining with fresh data?
Adding validation checks?
Improving your logging and alerting?
Simplifying the model or feature set?
Don’t just patch—prescribe. Build a fix that solves the root problem, not just the symptom.
4. Actioning: Operationalize the fix.
I used to think fixing the bug was enough.
But here’s what I learned the hard way:
If you don’t capture the fix, the same bug—or one just like it—will come back. Probably when you're on PTO.
Now, whenever I solve a tricky issue, I ask myself:
How do I make sure I never have to fight this fire again?
Here’s what I do:
✅ Jot down a quick checklist so I don’t miss it next time
✅ Add a test or alert to catch it earlier
✅ Write a short post-mortem—just for me or my team
✅ Share the fix in Slack, Notion, or a team doc
It only takes 10 minutes. But it turns a frustrating bug into a long-term system upgrade.
And more importantly?
It builds confidence—not just in your code, but in you.
The bonus step: Debug yourself.
This is what separates amateurs from pros.
Before you debug the model, take a minute to debug yourself:
Are you rushing because you're tired or under pressure?
Are you assuming too much?
Are your biases narrowing your thinking?
Clarity in systems starts with clarity in mindset.
Where this comes from
The NEPA framework originates from OPEX Fitness, an organization that trains elite fitness coaches to design customized, holistic training programs.
Their process: Notice → Explain → Prescribe → Act.
When I first saw it, I couldn’t help but think: “This is exactly how debugging should work.”
Borrowing mental models from outside tech can give you a serious edge. This one did for me—and I hope it will for you too.
Next time your model breaks:
Notice what changed.
Explain your hypotheses aloud.
Prescribe a targeted fix.
Action it into your workflow.
Debug yourself first.
PS… If follow me on LinkedIn, here’s what happens:
You’ll start seeing bite-sized frameworks, real-world debugging lessons, and practical mindset shifts—all tailored for early-career data scientists.
Why people stick around:
Because the content doesn’t just explain the what, it gives you the how—in a language that’s clear, no-fluff, and ready to use on Monday morning.
So, if you’d like a sneak peek into how I share frameworks like how I got rid of my impostor syndrome as a junior data scientist, click here to check out my latest post:
Until Next time,