17 y/o · Independent Researcher · India
Building artificial intelligence that goes beyond token prediction
I'm a self-taught researcher working without a CS degree, a lab, or a team. I reverse-engineer how the brain works to build AI systems fundamentally different from today's transformers.
My work spans neural architectures, training frameworks, cognitive systems, and — most recently — post-CMOS computing hardware. Different domains, same approach: rebuild from first principles, and eliminate what everyone else assumed was necessary.
A transistor-free computing architecture on hydrogen-passivated silicon. Computation happens by resonant electron absorption in 5-atom dangling-bond clusters, not by switching. On a 3 cm² die: 14.1 TB of in-situ memory at 79 mW — roughly 110× the M4 Max memory, at ~0.2% of the power. Memory and compute are the same atoms — no cache, no DRAM bus. My first paper outside machine learning.
Replaces standard attention with Q-Compass — sequence mixing grounded in reinforcement learning navigation theory instead of geometric similarity. Three projections instead of four. 69% fewer attention parameters. One mechanism handles text, vision, audio, and world modeling.
A modular multi-agent cognitive architecture featuring 12 specialized domain experts collaborating through Web-of-Thought (WoT) reasoning.
A plug-and-play fine-tuning framework that skips samples the model already knows — routing compute to hard samples and freezing mastered ones. Up to 80% compute savings at scale.
Berry-Q0 — ~50M parameters, trained from scratch on a single laptop GPU (RTX 4050, 6GB VRAM). Text + vision, currently in GRPO reasoning training (R1-style, math domain).
The goal: push a 50M model as far as possible on reasoning. No cloud. No team. Just architecture.