samvardhan03/omnipulse

OmniPulse is positioned as compliance and intellectual property infrastructure for the age of generative media. The basic idea is to take a piece of audio (the README focuses on 1 second clips of 44.1 kHz audio) and turn it into a mathematical fingerprint that can be stored, searched, and compared against a database, so that licensing and ownership claims can be checked quickly. The author describes performance numbers like fingerprint latency under 40 milliseconds on a Hopper-class GPU and sustained 15 GB per second host to device transfer. The project is a monorepo split into four modules that each ship to their respective package registries. omni-wst-core is a C++ and CUDA Python wheel that contains the digital signal processing math, based on Morlet wavelet scattering transforms. omni-ffi is a Rust crate that calls into the C++ side using the cxx crate, passing raw pointers so no data is marshalled or copied across the boundary. omnipulse-rs is a Rust workspace with a vector-index crate for HNSW nearest neighbor search, a sliced-wasserstein crate for comparing fingerprint distributions, and an MCP server binary called omnipulse-mcp that speaks line-delimited JSON-RPC 2.0 over stdio. omnipulse-agent is a Python package that uses an Anthropic-driven router to drive the whole pipeline. Data moves between the Python and Rust sides through POSIX shared memory, with a 28 character SHA3-256 hex name that fits the macOS shared memory name length limit. The audio buffer is written once and never re-serialized. The README emphasizes that there is no HTTP, no gRPC, no protobuf, and no TLS in this design. To try it, the README says to pip install omni-wst-core, cargo install omnipulse-mcp, pip install omnipulse-agent, set the ANTHROPIC_API_KEY environment variable, and run the agent against a WAV file. The license is Apache 2.0.