Kathleen: Oscillator-Based Byte-Level Text Classification Without Tokenization or Attention

Kathleen is a research-oriented project (0 stars, 1 fork, 8 days old) that attempts to bridge digital signal processing (DSP) with NLP. Its core innovation—using damped sinusoid convolutions (RecurrentOscillatorBanks) and FFT-based wavetable encoding—directly challenges the dominant Transformer paradigm for specific, resource-constrained classification tasks. From a competitive standpoint, the project has a low defensibility score (3) because it currently lacks adoption, community, and optimized kernels. While the math is novel, it is a 'white-box' algorithm that is easily reproducible once the paper's math is understood. It competes philosophically with State Space Models (SSMs) like Mamba or S4, which also offer O(L) complexity, but Kathleen is even more specialized for extreme parameter efficiency (733K parameters). Frontier risk is low because OpenAI and Anthropic are currently focused on scaling up, not scaling down to sub-1M parameter models. The platform domination risk is low because this is a niche architectural choice for edge devices or embedded systems rather than a general-purpose LLM capability. The main threat is displacement by more established 'efficient sequence' models like RWKV or Mamba, which have significantly more engineering momentum and optimized hardware kernels (e.g., CUDA/Triton support) that Kathleen currently lacks. If the oscillator-based approach scales well, it might be absorbed as a layer type in larger libraries (HuggingFace, etc.) rather than standing as a lone platform.