OmniPulse

OmniPulse is a media provenance platform for creators and rights organizations. It pairs two complementary verification modes on a shared substrate: OmniLock embeds a cryptographically signed 64-bit identifier into a video or image at creation (active, write-path), and OmniPulse computes a deterministic wavelet-scattering fingerprint that matches any derivative without cooperation at creation (passive, read-path). Both layers write to the same Ed25519-signed ledger and run on the same CUDA/Rust/Python substrate. The verify path is fixed-operator mathematics -- a linear parity check and an HNSW nearest-neighbor query -- so the verdict does not depend on a black-box service or on this team being available.

Live site: https://omnipulseid.vercel.app

Repository layout

omni-wst-core/          C++/CUDA DSP engine (WSTEngine, Morlet bank, Plasma shm)
omni-ffi/               cxx zero-copy FFI bridge (WST FFI family)
omnipulse-rs/           Rust workspace
  crates/
    omnipulse-mcp/      MCP binary (stdio JSON-RPC, HNSW, SW1, routing)
    vector-index/       concurrent HNSW
    sliced-wasserstein/ SW1 metric
    omni-lock-core/     OmniLock C-ABI v3, CUDA graph, build.rs gating
omnipulse-agent/        Python MCP control plane
site/                   Next.js marketing site (omnipulseid.vercel.app)
scripts/                FFI-separation and secrets CI checks

Two FFI families

There are two FFI seams and they never merge:

1. WST FFI (passive path): cxx::bridge between omni-ffi (Rust) and omni-wst-core (C++/CUDA). The bridge carries UVA-registered host pointers and raw u64 values; no JSON, no protobuf, no copy.

2. OmniLock FFI (active path): hand-written extern "C" ABI v3 in omni-lock-core/include/omnilock_ffi.h. Functions: omnilock_backend_create, omnilock_backend_destroy, omnilock_capture_inference_graph, omnilock_launch_inference_graph. Status codes: OK=0, ERR_NULL_PTR=-1 through ERR_NOT_INITIALIZED=-5. The two FFI families meet only in the Rust routing code of omnipulse-mcp, never in C++.

Why 28 characters for the shm name

hashlib.sha3_256(buf).digest()[:14].hex() produces a 28-character hex string. macOS caps POSIX shared-memory names at PSHMNAMLEN = 31 bytes including the leading /. 28 characters plus the prefix slash uses 29 bytes and fits every macOS and Linux kernel version without conditional logic. The name is a content digest, not a counter, so two identical buffers get the same name and the second shm_open returns the existing segment without an extra copy.

What is public and what is not

Rule: fixed operators are public; trained artifacts and secrets are private.

Public (this repo):

• All WST/JTFS kernels, Sliced-Wasserstein, HNSW, the cxx bridge
• omnipulse-agent, omnipulse-mcp, the site, all specs

Private (github.com/samvardhan03/omnipulse-engine):

• The full omni-lock-embed package: embedder, extractor, decoder, Mixer architecture, training pipeline, and all associated tests. These implement the write-path residual watermark and the trained Mixer; publishing them gives an adversary direct access to the embed procedure.
• The production LDPC parity-check matrix H and the generator that emits it. Publishing H turns the verify rule into a forgery target; the digest constant LDPC_H_SHA3_DIGEST is compiled into the public tree as a tamper check without revealing the matrix.
• Trained Mixer weights and the training procedure. The Mixer shapes the residual mask on the write path; publishing weights gives an adversary a warm start on a targeted attack.
• The Ed25519 issuer key and registry schema (secrets by definition).

The active embed/decode path hard-exits at build and runtime if OMNIPULSE_ENGINE_DIR is unset or does not contain the required artifacts. The passive fingerprint path has no such gate and builds from this repo alone.

The verify path (passive fingerprint, active parity check + signature) uses only fixed operators and is fully inspectable in this repo. The write path uses trained shaping that stays private. This is a coherent split: anyone can audit the verification rule; the embedding quality is our moat.

Quickstart: passive fingerprint path (no private artifacts needed)

The passive path -- fingerprint a media file and insert it into the HNSW index -- builds and runs from this repo alone. The active embed/decode path requires engine artifacts from the private repo; it fails loudly with instructions if those artifacts are absent (see omni-lock-core/build.rs).

Requirements (passive path)

• Python 3.11+
• Rust 1.78+ (rustup show)
• C++17 compiler
• CUDA Toolkit 11.8+ for the GPU path; the CPU Morlet fallback (--features omni-ffi) works without CUDA

Steps

git clone https://github.com/samvardhan03/Omnipulse.git
cd Omnipulse

# 1. Python packages
pip install omni-wst-core omnipulse-agent

# 2. Rust MCP orchestrator (passive path, CPU Morlet fallback if no CUDA)
cargo build -p omnipulse-mcp --features omni-ffi

# 3. Fingerprint a WAV file
export ANTHROPIC_API_KEY=sk-ant-...
python -m omnipulse_agent.run --wav your.wav

You will see the four-stage trace: ingest, shm pin (28-char SHA3 name printed), cxx bridge, HNSW insert.

Status

License

AGPL-3.0 for open-source and research use. Commercial license for production deployments that cannot comply with AGPL source-disclosure requirements. Contact [email protected] for commercial terms.

See LICENSING.md for the dual-licensing model. See LICENSE for the full AGPL-3.0 text.

Maintainers

Samvardhan Singh -- systems, signal processing, CUDA substrate. samvardhan.vercel.app / [email protected]

Yash Mishra -- scattering research and the passive engine. linkedin.com/in/mishra-yash2002

Shreyansh Jain -- agentic systems and the active layer (OmniLock). shreyanshjain05.vercel.app / github.com/shreyanshjain05

Contributing

See CONTRIBUTING.md.