Media provenance and rights infrastructure

Provenance that lives inside the pixels.

A signed identifier embedded in your media at creation, a fingerprint that finds every derivative after, and a record that belongs to you.

See it work Talk to us

The identifier survives what the internet does to media.

Problem

A pitch-shifted copy walks straight past waveform matching.

What it is

The identifier lives inside the media, not in a sidecar a re-upload can strip. Verification is deterministic mathematics, not a black-box service.

Why it is different

Two engines, one substrate: OmniLock writes a 64-bit LDPC watermark. OmniPulse reads WST/JTFS scattering coefficients. Neither requires a cloud call to verify.

Who it is for

Independent creators, studios, and rights organizations who need to license, track, and monetize their work without ceding the record to a platform.

Module stack

omni-wst-core

0.1PyPI

omni-ffi

0.1crate

omni-hnsw

0.1crate

omni-sw

0.1crate

omnipulse-agent

0.1PyPI

One record, two ways to reach it

Two engines. One registry. One signed token.

Active layer

OmniLock

At creation, we embed a 64-bit signed identifier into the video's motion-compensated residual, in the frequency band a modern codec preserves and the eye ignores. Verification recovers your identifier exactly, or reports that it cannot. No probabilities.

DCT band explorer

Hover or tab to any cell to see its role.

25 of 64 coefficients per block: chosen by codec physics, not by preference.

Passive layer

OmniPulse

When there was no cooperation at creation, there is nothing to extract. We compute a deterministic wavelet-scattering fingerprint and match it against your catalogue. Same registry, same signed token, different physics.

Attack survivability

waveform matching (legacy)92%

OmniPulse (JTFS)94%

Illustrative demo, not a live benchmark.

Talk to us

OmniLock embed simulator

Embed a signed identifier. Verify it back.

Choose a sample image or drop your own. The 64-bit identifier is derived deterministically from the image bytes, in your browser.

1. Ingest

2. DCT transform

3. LDPC encode

4. Embed

5. Token minted

Drop or click to upload your own imageProcessed entirely in your browser, never uploaded

Pipeline

ID derived from SHA3-256 of image bytes, truncated to 64 bits.

LDPC expands 64 bits to 128-bit codeword. Parity bits shown in orange.

Mask is confined to DCT mid-band (2 <= u+v <= 6) per codec physics.

Token is an Ed25519 signature over ID + timestamp + IPFS CID.

Demo runs entirely in your browser with a demonstration key. Production embedding runs on our GPU backend.

Talk to us

Honest comparison

SynthID and OmniPulse answer different questions.

SynthID answers a different question. Both matter.

Dimension	SynthID (Google DeepMind)	OmniPulse
Who owns the record	The provider	The creator, on a decentralized ledger
What it verifies	Whether content was AI-generated	Whether content is yours, and under what license
How verification works	Through the provider's service	Deterministic mathematics anyone can run

Talk to us

Modular commercialisation

Four tiers, one fingerprint plane

Phase I

Phase I: Institutional Research

Standalone DSP wheels for desktop labs and HPC clusters. C++/CUDA Wavelet Scattering primitives, no orchestration required. AGPL-3.0 + Commercial.

Scope:Passive audio only.

pip install omni-wst-core

→ Know how it worksView on PyPI

Phase II

Phase II: High-Velocity Concurrent Processing

The vector substrate. omni-hnsw + sliced-wasserstein Rust crates for sub-millisecond nearest-neighbour over billion-scale fingerprint catalogues. AGPL-3.0 + Commercial.

Scope:Passive across all modalities, small catalogue.

cargo add vector-index sliced-wasserstein

→ Know how it worksView on crates.io

Phase III

Phase III: Autonomous Agentic Control

Drop-in MCP server. omnipulse-agent parses operator requests and routes them to the Rust orchestrator over line-delimited JSON-RPC 2.0. AGPL-3.0 + Commercial.

Scope:Active embed for images and video, full catalogue.

pip install omnipulse-agent

→ Know how it worksView on PyPI

Phase ENTERPRISE

Enterprise SaaS Pipeline

Commercial

Cloud-native multi-tenant deployment. Kubernetes control plane, GPU autoscaling driven by Prometheus queue depth, ed25519-signed license tokens, IPFS pinning.

Scope:Dedicated instance, versioned verification procedures.

→ Know how it worksTalk to founders

Interactive simulator

The fingerprint pipeline, step by step

Drop a WAV file. Step through each swimlane. Every value is deterministically computed from your input, no server calls.

Upload audio

Tensor ingested

SHM pinned

RPC in-flight

FFI bridge

Kernel running

Completed

Drop or click to upload a WAV fileDecoded in-browser via AudioContext · no upload

Python control plane

# SharedMemoryManager.ingest_media_tensor audio = np.array([...], dtype=np.float32) shm_name = shm.ingest_media_tensor(audio) # → "..." (28 hex chars)

Shared-memory pointer

/a3f7b2c1d4e5f6a7b8c9d0e1

SHA3-256(raw).digest()[:14].hex() - 28 chars fits macOS PSHMNAMLEN=31

Stdio JSON-RPC envelope

{
  "jsonrpc": "2.0",
  "id": "01J4...",
  "method": "tools/call",
  "params": {
    "name": "generate_fingerprint",
    "arguments": { ... }
  }
}

Rust orchestrator

// server.rs:162-211 - generate_fingerprint handler let signal = shm::read_and_unlink(&req.media_shm_name, n * 4)?; let fp = tokio::task::spawn_blocking(move || { kernel.run(&signal, &cfg) }).await??;

cxx FFI bridge

// omni_ffi_kernel.rs:90-97 unsafe fn run_wst_pipeline( ptr: u64, len: u32, j: u32, q: u32, depth: u32 ) → WSTResult // → omni_ffi::execute_fingerprint_pass(...)

C++/CUDA kernel

// CPU Morlet fallback - wst_bridge_cpu.cpp AnalyticMorletBank<8, 16> bank; bank.scatter_radix2(ptr, len); // Radix-2 FFT # build: cargo build -p omnipulse-mcp --features omni-ffi

HNSW index (SW1 metric)

ConcurrentHnsw<PointCloud, SlicedWasserstein> .insert(fp, id) // node 0 // SW₁ distance < 1e-6 → nearest neighbour

WstConfig

J = 8

Q = 16

JTFS

Kernel

CPU Morlet (Radix-2 FFT)

Custom Pricing & Enterprise Use Cases

Built for organizations where provenance is a legal question.

One fingerprint plane, three high-stakes domains. Every deployment is on-prem, GPU-accelerated, and produces a write-once audit trail. Pricing is contract-based; talk to us.

Major Label · Studio

Derivative detection at catalogue scale

Detect unauthorized derivatives, interpolations, and AI reproductions across a 100 M-asset library. Deterministic Morlet scattering fingerprints (no model, no retraining, no false-positive amnesty). Under 40 ms per track on H100.

How:Fingerprint the catalogue at ingest (passive), embed new releases at creation (active), one signed ledger.

under 40 ms · Apache audit log · on-prem, no egress

Generative Platforms (Suno, Udio)

IP verification before distribution

Every generation gets a fingerprint at creation time. Match probability is computed against the rights-holder catalogue before the track ships. The JSON-RPC audit trail writes the compliance documentation automatically.

How:Sign outputs at generation; downstream verification never depends on you being online.

per-generation attestation · IPFS pin hash · licensable origin certificate

IP Litigation · Forensics

Expert-witness-grade match probability

The Sliced-Wasserstein distance between two fingerprints is a number a judge can examine. Every comparison frame is a write-once JSON line. No opaque model, no black-box score: a mathematically interpretable result.

How:A linear parity check and an Ed25519 signature: public procedures, bit-exactly reproducible.

full audit replay · on-prem · zero data egress

Custom PricingContract-based · SLA committed

Talk to founders →

Built on the same engine

Two sister products. Same data plane. Different physics.

The OmniPulse scattering engine (omni-wst-core, omni-ffi, sliced-wasserstein, vector-index) also runs at the heart of two independent product surfaces. Same zero-copy core, same mathematical guarantees, completely different markets.

SCIENTIFIC COMPUTE · vikshep.vercel.app

ψ Vikshep

Deterministic features for physics. Nothing learned, nothing leaked.

Wavelet scattering features for physics analyses: mass-decorrelated jet tagging, template-free BSM anomaly detection, rotation-invariant field inference. Because the filters are fixed by the mathematics of the scattering transform and not by training, the features cannot secretly encode the resonance mass, so a downstream cut cannot sculpt a fake bump into a smooth background. Currently in pilot with the University of Edinburgh on Geant4-simulated ATLAS diboson data.

· 1-D / 2-D / 3-D runtime config· SE(2) · SO(3) invariance· r₂ + DisCo decorrelation

Visit Vikshep →

SYSTEMATIC ALPHA · drift-site-livid.vercel.app

Drift

From log-returns to alpha. Seven layers. Zero black boxes.

Quantitative research platform built on the same scattering core. Cauchy wavelet features over log-returns, rank-transform IC/ICIR signal composition, Black-Litterman and HRP portfolio construction. Every formula is implemented directly from first principles (no hidden models, every step derivable from the build history). 182 tests, 0.70 Sharpe on the 3-year synthetic HRP backtest, MCP-native.

· 7 layers from data to weights· 182 tests passing· Sharpe 0.70 · PSR 0.886

Visit Drift →

Three product surfaces. One engine, dual-licensed AGPL-3.0 + Commercial. Maintained by the same three people.

Founders & creators

The people behind the platform

ARCHITECT, APPLIED AI / MLOPS

Samvardhan Singh

Systems, signal processing, and the CUDA substrate

Automation engineering, AI/MLOps pipelines, engineering outcomes.

samvardhan.vercel.app/github.com/samvardhan03 email: shekhawatsamvardhan@gmail.com

Maintains

omni-wst-core- C++/CUDA DSP engine omni-ffi- cxx zero-copy FFI bridge omnipulse-agent- Python agentic control plane (PyPI)

ARCHITECT, SYSTEMS / OPTIMAL TRANSPORT

Yash Mishra

Scattering research and the passive engine

Concurrent systems, optimal transport, real-time indexing logic.

www.linkedin.com/in/mishra-yash2002/email: yash01012002@gmail.com

Maintains

vector-index- Concurrent HNSW sliced-wasserstein- SW1 distance metric for HNSW

ENGINEER, AGENTIC SYSTEMS / ACTIVE LAYER

Shreyansh Jain

Agentic systems and the active layer (OmniLock)

GenAI and agentic-systems engineer; research publications and open-source Python packages.

shreyanshjain05.vercel.app www.linkedin.com/in/shreyanshjain05/github.com/shreyanshjain05

Maintains

omni-lock-core- CUDA inference crate (C-ABI v3)omni-lock-embed- PyTorch embedder

Co-authored: Phase 3: The Agentic Control Plane is jointly maintained by Samvardhan and Yash across omnipulse-agent (Python) and the omnipulse-mcp server in omnipulse-rs.

Pilot program

Run a pilot with us.

We are onboarding a small number of pilot partners for the merged platform: labels, generative platforms, and rights organizations. You bring a catalogue or an output stream; we bring the registry, the embedding, and the verification procedure. Honest engineering, no black boxes, and we will tell you what is measured versus what is still a budget.

Start the conversation

Or reach us directly at shekhawatsamvardhan@gmail.com

What a pilot includes

Registry setup and key provisioning

Catalogue fingerprinting run

Active embed test on a sample release

Verification procedure walkthrough

Honest report on what worked and what did not

Get started

Ready to integrate the fingerprint plane?

Start with the open-source modules. The core is AGPL-3.0 + Commercial; enterprise production is a conversation.

View the Showcase Repo →