A cognitive processing system implementing neural architecture search, hierarchical state spaces, mixture of experts, and a full S0–S9 security layer — built from the ground up for secure, aligned, performant AI.
SneppX-ALG is a full-stack cognitive processing system designed to bridge the gap between low-level hardware acceleration and high-level neural computation. It implements neural architecture search (NAS), hierarchical state spaces (HSS), mixture of experts (MoE), and a comprehensive S0–S9 security framework within a single, unified architecture.
Written in C11 and C++20 with CUDA acceleration, targeting x86-64 with Python bindings via pybind11. Every component — from the tensor engine to the distributed training runtime — is designed for audibility, determinism, and asymptotic efficiency. Security is structural: cryptographic primitives, memory hardening, formal verification, and AI safety are compiled directly into the foundation layer.
Core tensor engine and automatic differentiation engine built from scratch in C11.
Cryptographic primitives, memory hardening, obfuscation, runtime monitoring, and network security implemented.
LLaMA, Mistral, and Qwen2 support with INT8/FP8 quantization pipelines.
ZeRO-1/2/3, pipeline, tensor, and expert parallelism across multi-node clusters.
Flash Attention v2/v3, fused GEMM kernels, NCCL all-reduce, and memory pool.
Continuous batching, quantized serving engine, and full CLI toolchain.
Security is not an afterthought; it's woven into every layer — from hardware primitives to AI safety.
Hardware-accelerated from the ground up — CUDA, AVX-512, AES-NI, SHA-NI, and tensor cores.
Clean Python API, comprehensive documentation, CLI tools, and fast iteration cycles.
Open source, community-driven, MIT licensed. Built in public on GitHub.