**Requirements:**
- 2+ years in digital marketing/coordination
- CMS, social media management
- Excellent English
- M365/Teams proficiency

**Bonus:**
- Healthcare/education/NPO experience
- Cvent/CRM/SEO skills
- EU languages

👉 Apply: careers.envistaco.com/job/2261483...

Full analysis with code and methodology: github.com/Tuminha/Per...
And full story here: substack.com/home/post/p...
Built this for my @Periospot dental education platform but the lessons apply everywhere.

Final thought: I spent months trying to engineer features to predict viral posts.
The model taught me I was asking the wrong question.
Not "what makes posts viral?" but "why do we overestimate content's impact?"
Distribution > Creation. Every time.

For fellow data scientists: This is why temporal validation matters. Train on 2024, test on 2025 revealed the Instagram algorithm shift. Random split would've hidden it.
Always test on future data for social media models.

The model's "failure" was its success. ROC-AUC 0.62 isn't bad at prediction - it's honest about reality.
Sometimes ML's greatest value isn't in what it predicts, but in what it proves isn't predictable.

What this actually means:
Stop optimizing:

Perfect posting times
Ideal hashtag counts
Caption length formulas

Start optimizing:

Follower growth
Platform selection
Consistency over perfection

This reminded me of Sutskever's detective analogy for LLMs - "predict the killer from the clues."
But in social media, the killer was decided before entering the room. The platform and follower count predetermined 70% of the outcome.

youtu.be/c2UGZmmgd0g

Platform-specific performance:
X/Twitter: Model caught 74% of high performers
Instagram: 0% (literally none to catch)
Threads: 30% recall
Same model. Same approach. Completely different results.

The Instagram collapse:
2024: 3-5% average engagement
2025: 0.2% average engagement
Not a single Instagram post in 2025 qualified as "high performing" by my model's standards. The platform changed. The content didn't.

Feature importance was depressing:

has_numbers: 11%
video vs photo: 10%
hashtag_count: 6%
caption_length: 5%
posting_hour: 4%

Even the "best" content features barely matter.

The math is brutal:
~70% of engagement variance = which account/platform (distribution)
~30% = actual content quality
Your follower count matters more than your writing. By a lot.

First red flag: When I included account/platform features, the model hit 0.77 ROC-AUC.
When I removed them to focus on pure content? Dropped to 0.62.
That 20% drop = the entire game.

Setup: I had posts from X, Instagram, Threads, LinkedIn, Facebook, YouTube. All my @Periospot content.
Built an XGBoost model with every feature I could think of:

Posting time
Hashtags
Caption length
Emojis
Content type

Goal: Predict high-performing posts