an ai powered hospital bill auditor, best mobile hack at hacktj. photo/pdf upload, extracts medical codes and pricing, validates against a cms database via rag, flags fraudulent charges, and drafts a formal dispute letter.
Anay Shekhar
15 y/o learning ML
about
i'm someone who just wants to know what's actually happening underneath the model, not just that it works. that curiosity is what pulled me into machine learning in the first place, and it's why i build things from scratch instead of importing them. if i can't explain why a gradient flows the way it does or why a representation looks the way it does, i don't feel like i understand it yet.
right now that curiosity is pointed at world models and interpretability, specifically how models learn to represent physical reality instead of just predicting text, and what's actually going on inside them when they do.
now
- joint embedding predictive architectures (jepa)
- latent space prediction vs pixel reconstruction
- probing what these models represent internally
projects
a language model built entirely from scratch in numpy that generates pokémon names. no frameworks, manual forward and backward passes, built to understand weight initialization, activation stability, and gradient flow directly.
future
looking ahead, i want to keep asking whether models build real internal representations of the world, or just get good at faking it. i'd rather spend time figuring out what's actually happening inside a system than chase whatever's trending.
- representation learning: building a small jepa style model from scratch to see how a context encoder and a target encoder learn to predict masked regions in latent space
- mechanistic interpretability: writing probes to check what a model's latent space actually encodes, like object permanence or spatial structure, instead of assuming it does
- evaluation: studying where current world models fail on physical reasoning benchmarks, and what that reveals about the gap between representation and understanding