Engineering for the Long Game
In The Loop: Building Hard Things, Learning Harder
If you asked me how it feels after two startup exits and shipping at two tech giants — honestly, it feels like just another day.
Two exits: Snowflake going public, and OctoML being acquired by NVIDIA. When Snowflake went public, I didn’t feel much. I went downstairs, grabbed a coffee, and went back to building — on my day off. That’s just the rhythm: stay grounded, stay focused, keep building.
OctoML was different. That one felt personal. We’d been building for years — pivoting, iterating, hitting walls. There were hard times and long stretches where it wasn’t clear what would work. But there was solidarity. I’m proud of what we built together — from the PyTorch-based inference service and TensorRT-LLM compilation work to seeing the team’s ideas live on at NVIDIA. And I’m even more proud of the people — Luis, Tianqi, Jared, Jason, Itay, and so many others who made it a memorable, human experience. You don’t forget those kinds of teams.
- Between Giants
Then came two tech giants: NVIDIA and Apple. They couldn’t be more different — yet both demand excellence in their own way.
NVIDIA runs lean and focused. Every decision has a purpose. You learn what “execution speed” really means when the bar is that high. Apple, on the other hand, is deliberate and design-driven — but I believe it’s perfectly positioned to ship the next generation of AI products.
Since joining Apple, I’ve had the chance to work on Apple’s foundation models, and contribute to Apple Intelligence. Seeing things go from research to shipping products — that still hits deep. It’s hard, technical work, but it matters.
Building, Breaking, and Learning Again
- The Hard Way is the Only Way
If I had to choose between vision, idea, and execution, I’d pick vision.
Every Single Time
At NVIDIA there’s a term — Speed of Light (SOL). It doesn’t mean deliver the fastest; it means understand the right problem before you start running. Execution without a strategy is just noise. Low-hanging fruit feels good in the short term but leads to mediocrity in the long run. The order matters:
Find the hardest problem → Build the right strategy → Execute fast.
- Learn or Leave
Learning has been the one constant. If our DNA is what’s been pre-trained for millions of years, then learning is post-training.
Don’t leave when things get tough — leave when you’ve stopped learning. That’s the real warning sign. Learning is painful, but not learning is worse.
- Work with People You Respect
It’s a two-way relationship. Surround yourself with people you want to work with — people who stay curious, who argue with you for the right reasons, who keep ego low and know better about ownership than being territorial.
Every time I joined a new place for the last few years, I’ve been constantly amazed by the talent density and focus of the people who truly care about their craft. Those are the people worth learning from.
Next Chapter
These days, I’m working on post-training & optimizations on foundation models at Apple. It’s one of the hardest and most important problems in AI right now. I believe post-training — how we align and adapt models — will be key to understand what intelligence really is. But it won’t just scale; it’ll take new infrastructure, new data, and new ideas to get there.
A couple thoughts:
- Multi-modality — to learn richer representations and bridge data from the world.
- Agents — to push AI toward true autonomy and usefulness.
- Embodiment — to close the loop between intelligence and reality.
And maybe one last thought — the people who care deeply about software often end up building hardware. The people who truly believe in AI are usually the ones who care most about making it safe.
That balance — ambition and responsibility — is what keeps me going, for the long game.