Building novel neural architectures grounded in physics. Created Resonance, where the damped harmonic oscillator transfer function H(ω) = 1/(ω₀² - ω² + 2iγω) is the only learnable transform. Discovered that signal-native encoding — feeding raw data directly into the oscillator bank — outperforms hand-designed encodings by 4x. Architecture scales consistently (PPL 1.77 at 21K to 1.14 at 1M params), works for both audio and text, and is provably an energy-based model. Previously built Vidya (neurosymbolic LM in OCaml) and sshmail (encrypted messaging over SSH in Go). All work open-source.