Want to know why some ML projects soar while others crash and burn? After 6 years in the ML trenches at AWS and now Nebius, Alex Pathrushev has seen it all. Here’s the real deal on what makes or breaks ML projects in the real world.
About the Speakers
Alex Pathrushev
VP of AI/ML at Nebius, Alex brings over 6 years of deep ML expertise from leadership roles at AWS and Nebius. At Nebius, he’s driving the development of next-generation cloud infrastructure specifically optimized for AI workloads. His team has built a full-stack cloud platform that’s making waves in the ML space, with data centers across Europe and the US.
About Nebius
Nebius is reshaping how companies build and deploy AI, with a team of 500+ engineers across Europe and the US. Their platform combines custom-built hardware, state-of-the-art data centers, and an advanced cloud platform designed specifically for AI workloads. With R&D teams actively contributing to open-source ML models, Nebius is positioning itself at the forefront of accessible enterprise AI.
The TL;DR: Your 5 Must-Have Pillars
- Own Your Data (Like Really Own It)
- Build a Dream Team (But Make It Real)
- Master Stakeholder Communication
- Pick Winners (And Know When to Kill Projects)
- Get Smart About Your Tech Stack
Let’s dive deep into each one.
1. Own Your Data (Because Garbage In = Garbage Out)
Here’s the brutal truth: Without solid data, you’re dead in the water. Period.
Before you write a single line of code, answer these three questions:
- What data do you have RIGHT NOW?
- What data could you grab quickly?
- What data should you start collecting TODAY for tomorrow?
The secret sauce? Your data needs three key ingredients:
Availability
You need it when you need it. Sounds obvious, right? But you’d be shocked how many projects start with “we’ll get the data later.” Spoiler alert: They usually fail.
Quality
This isn’t just about clean data (though that’s crucial). It’s about having the RIGHT data. We’ve seen teams do better with less data that’s high quality than mountains of junk data.
Diversity
Your model needs to handle the real world, not just your test environment. Diverse data = robust models.
Real Talk: We saw this firsthand building a code generation system. The team thought they had great data until they realized they were missing crucial context about execution environments and results. After finding and cleaning the right open-source data, they hit a 40.6% benchmark – without using any commercial models.
2. Build Your Dream Team (But Keep It Real)
The minimum viable ML team needs three key players:
- Data Scientist: Your ML expert
- Data Engineer: Your data pipeline guru
- MLOps Engineer: Your production deployment master
Don’t have all these folks? No problem. Get creative:
- Partner up with cloud providers
- Tap into academies
- Bring in external experts
Success Story: Recraft built one of the best design tools in the market by nailing this collaboration piece. They didn’t just throw people together – they built a team that could solve problems fast and iterate even faster.
3. Master Stakeholder Communication
Here’s where most teams drop the ball: They treat ML projects like pure tech plays. Big mistake.
The winning formula:
- Open lines of communication (ALL the time, not just when things go wrong)
- Crystal clear expectations from day one
- Direct connection between ML work and business outcomes
Pro Tip: Start every project by answering: “How will this make us money, make customers happier, or bring something new to the table?”
4. Pick Winners (And Know When to Walk Away)
Here’s the framework that’s saved us countless hours and dollars. Score every potential project 1-9 on:
- Data Availability
- Business Impact
- ML Feasibility
The secret? You need balance. A perfect score in business impact means nothing if you have no data or if the ML piece is a moonshot.
5. Get Smart About Your Tech Stack
Here’s the golden rule: Don’t rebuild what you can reuse.
The smart approach:
- Survey what’s out there (open source and commercial)
- Pick what works NOW
- Optimize LATER (after you’ve proven value)
Remember: Your job is to drive business results, not build the world’s most elegant tech stack.
What’s Next? The Future of ML in SaaS
The next big wave? Genetic systems that can:
- Write their own code
- Make autonomous decisions
- Adapt to changing conditions in real-time
We’re seeing this already in:
- Enhanced shopping experiences
- Automated customer service
- Dynamic workflow optimization
The Bottom Line
Success in ML isn’t about having the biggest budget or the smartest algorithms. It’s about nailing these five fundamentals and executing relentlessly.
Get these right, and you’re already ahead of 80% of ML projects out there. Miss any one of them, and you’re in for a world of pain.
Want to dive deeper? Join the Nebius Discord channel or catch us at the next SaaStr event. We’re always happy to geek out about ML done right.
