4 things it takes to build an incredible company:
Passion.
Discipline.
Obsession.
And the most important: Flamingo. ๐ฆฉ
Building Quettos, the company that powers intelligence, takes all four. Letโs go win.
I'm building the next nvidia:
Checklist:
- Engineering leather jacket check
- Engineering iron ring check
- Turtle neck for added aura check
- Inevitable victory mindset check
Alright let's go rock the world.
@r4f_newsletter Great question! Weโre co-designing our compiler and hardware for agent workloads. TileLang is a good development for the whole ecosystem. At Quettos we want high performance to be the default, not something developers have to fight for.
Alright, let's get it!
A few months ago, I founded Quettos with one mission: build the next generation of computing systems to run intelligence. To achieve this goal, we are starting with AI inference chips purpose-built for agent workloads.
As of November, we've grown into a founding team of three. We are moving fast and having fun tackling some of the hardest challenges in computer architecture.
Currently building out of the Velocity incubator in Waterloo.
Lots to build. Lots to do.
And we're just getting started.
One of my favorite things about Rust: No garbage collector.
No manual memory management.
Ownership and borrowing let the compiler prove who owns what at compile time.
It's a really clean and elegant idea.
Ever sorted an array? Try sorting one in hardware.
No software loop iterating through the array one comparison at a time.
Just comparators wired between neighbors: each clock cycle, every adjacent pair gets compared in parallel and the out-of-order ones swap. The array falls into place over a handful of cycles.
This is an odd-even transposition sort I built in university: one compare-and-swap stage, reused each clock edge.
Software sorts in time. Hardware sorts in space.
Lost ~$40 because I forgot I'd disabled flash-attention.
Left an H200 training overnight, came back to find every step running ~4x slow on the fallback attention kernel.
Long context + no flash-attn = quadratic tax.
Lesson: don't touch the training config when you're tired.
During my last term at the University of Waterloo, I led my capstone team in the ambitious project of building TurnSight AI, a smart left-turn assist system.
Here's the system in action. While driving, you can see our embedded system hard at work, and in the bottom right you can see exactly what the system sees in real time.
TurnSight watches oncoming traffic and warns new drivers when it isn't safe to turn, basically acting as an extra set of eyes on the dashboard.
The whole thing runs locally on an NVIDIA Jetson Orin Nano. No server, no cloud, just edge AI.
The hardest part wasn't getting an AI model to detect cars. It was getting everything fast enough for real-time inference on embedded hardware. A lot of teams at the symposium had AI systems too, but many couldn't actually run in real time.
Making YOLO11 keep up with live video on the Orin was a grind. I rebuilt OpenCV from source for CUDA support, converted the model to TensorRT, and tuned the pipeline end to end until it could actually run in real time.
On top of that, we built custom threat detection logic to determine whether an oncoming vehicle is actually a threat.
One of the professors who stopped by our booth told us we should patent it.
Still letting that one sink in.
Anyway, that's one chapter of the story.
On to the next!
I'm building a chip company by 30. Before I dive in, a quick trip down memory lane felt overdue.
While most courses at UWaterloo were pretty dull, a few really stuck with me. One of them was Computer Architecture where I built my own CPU!
Back then, I spent way too many nights in the lab room staring at screens like this, trying to figure out why my code was breaking in eight different ways at once.
Now I'm starting down the road to building chips for future AI workloads!