The advice you're getting is practical, but it's missing a layer. LeetCode alone won't prepare you for the applied scientist or research engineer interviews you're targeting. Those roles live and die by your ability to think in tensors, reason about memory layouts, and write clean PyTorch under pressure. The good news is you've already found the right tools. The challenge is choosing where to spend your limited hours, and that choice should hinge on one question: what exactly are you trying to simulate?
If your goal is to feel comfortable with the kind of low-level, implementation-heavy questions that come up in applied ML interviews, then deep-ml and tensorgym are your best bets. They push you to implement layers, loss functions, and training loops from scratch, which is exactly the muscle memory you need. LeetGPU leans more toward CUDA and performance engineering, which is valuable if you're aiming for a role that touches kernels or inference optimization, but it's a narrower slice of the pie. NeetCode's PyTorch section is fine for brushing up on common patterns, but it's not where you'll build fluency. And nexskillai is worth a look, but you should treat it as a supplement, not a primary source, until you've confirmed it matches the depth of the others.
Here's the practical move: pick one platform and commit to it for two weeks. Do a few problems every day, but more importantly, write down the reasoning behind each line you write. Why did you use `view` instead of `reshape`? When would you reach for `torch.where` over a mask? The platforms you listed will all give you reps, but only you can turn those reps into a mental model. And that model is what separates candidates who memorize APIs from those who can reason about numerical stability or gradient flow on the spot.
One more thing: don't drop LeetCode entirely, but shift your ratio. Spend 70% of your prep time on PyTorch and NumPy, and keep the remaining 30% for algorithmic fundamentals. That mirrors what you'll actually face in interviews, where you'll get a mix of classic coding and applied ML questions in the same session. Start with tensorgym or deep-ml, whichever feels more uncomfortable, and build from there. The goal isn't to try every tool. It's to make one of them feel like second nature before you walk into the room.