Omnipulse
Health Uyari
- No license — Repository has no license file
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Community trust — 20 GitHub stars
Code Gecti
- Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
- Permissions — No dangerous permissions requested
Bu listing icin henuz AI raporu yok.
Enterprise IP infrastructure for the generative media era. A polyglot (C++/Rust/Python) zero-copy perceptual hashing and differential licensing plane using Wavelet Scattering Transforms and MCP-native agentic orchestration.
Ω OmniPulse
Compliance & IP Infrastructure for the Generative Media Era
Wavelet-scattering fingerprints. Zero-copy FFI. Memory-safe orchestration. MCP-native.
This repository is the complete OmniPulse monorepo.
All four modules live here and ship to PyPI and crates.io independently.
This root lets you read the architecture in one place and jump to
whichever sub-package you need. See https://github.com/samvardhan03/Omnipulse.
Engineering outcomes (the headline numbers)
| Property | Outcome | Where it comes from |
|---|---|---|
| End-to-end fingerprint latency | < 40 ms for a 1-second 44.1 kHz audio clip on a single Hopper-class GPU | C++/CUDA WSTEngine<HopperTag, J, Q> on pre-pinned UVA pages |
| Host ↔ device transfer throughput | 15+ GB/s sustained DMA on PCIe 4.0 x16 via Apache Arrow Plasma pages registered with cudaHostRegister |
cpp/wst_bridge.cu — pinned, portable, no cudaMemcpy round-trip |
| Python ↔ Rust handover cost | Zero-copy — the f32 audio buffer is written once to POSIX shared memory and never re-serialized | omnipulse-agent → omnipulse-rs over a 28-char SHA3-256 hex shm name |
| FFI boundary cost | Zero-marshalling — cxx passes raw u64 pointers across the C++↔Rust seam |
omni-ffi — unsafe extern "C++" fn run_wst_pipeline(input_plasma_ptr: u64, ...) |
| Memory safety | RAII-guarded WSTResult ownership; shm_unlink is single-owner and idempotent |
OmniFfiKernel::WstResultGuard in omnipulse-rs; shm::read_and_unlink |
| Concurrency | Lock-shared HNSW reads + bounded-time writes under parking_lot::RwLock; blocking kernel runs on tokio::task::spawn_blocking so stdio stays live |
ConcurrentHnsw<PointCloud, SlicedWasserstein> |
| Transport overhead | Line-delimited JSON-RPC 2.0 over stdio — no HTTP, no TLS, no socket setup | rmcp transport::stdio() |
| OS portability of segment names | 28-char shm names fit the macOS PSHMNAMLEN = 31 ceiling and every Linux kernel ever shipped |
hashlib.sha3_256(buf).digest()[:14].hex() |
These are the operating points the architecture optimizes for. The
underlying math (analytic Morlet wavelet scattering, sliced
1-Wasserstein on empirical fingerprint distributions, HNSW indexing) is
explained in depth inside the per-module READMEs — this root keeps
the lens on engineering outcomes.
Architecture at a glance
┌──────────────────────────────────────────────┐
│ omnipulse-agent (PyPI · Python 3.11+) │
│ ────────────────────────────────────── │
│ • Anthropic-driven cognitive router │
│ • SharedMemoryManager (POSIX shm) │
│ • MCPClient — line-delimited JSON-RPC 2.0 │
└────────────────────┬─────────────────────────┘
│
28-char hex shm name │ stdio pipes
(SHA3-256 digest[:14]) │ (one \n-terminated frame per call)
▼
┌──────────────────────────────────────────────┐
│ omnipulse-rs (crates.io · Rust workspace) │
│ ────────────────────────────────────── │
│ • omnipulse-mcp binary (rmcp stdio) │
│ • vector-index (concurrent HNSW) │
│ • sliced-wasserstein (SW₁ metric) │
│ • shm_open + mmap + spawn_blocking │
└────────────────────┬─────────────────────────┘
│
u64 pinned-host ptr │ cxx::bridge (zero-marshalling)
(UVA-registered page) │
▼
┌──────────────────────────────────────────────┐
│ omni-ffi (crates.io · zero-copy bridge) │
│ ────────────────────────────────────── │
│ • cxx 1.0 — unsafe extern "C++" │
│ • CPU path: cpp/wst_bridge_cpu.cpp │
│ • CUDA path: cpp/wst_bridge_cuda.cpp │
│ (links cudart + cufft) │
└────────────────────┬─────────────────────────┘
│
raw float* / CUdeviceptr│
▼
┌──────────────────────────────────────────────┐
│ omni-wst-core (PyPI · C++/CUDA wheels) │
│ ────────────────────────────────────── │
│ • WSTEngine<HopperTag, J, Q> (CUDA) │
│ • build_cpu_morlet_bank + Radix-2 FFT (CPU) │
│ • Plasma cudaHostRegister │
└──────────────────────────────────────────────┘
Every arrow is one of: a 28-char shm name, a \n-terminated JSON
frame, a raw u64 pointer, or a registered host page. No protobuf,
no gRPC, no HTTP, no JSON-over-network. That is the design.
Modules — where each package lives
| Layer | Repo | Package | Quickstart |
|---|---|---|---|
| DSP primitives (C++/CUDA) | samvardhan03/Omnipulse — omni-wst-core/ |
omni-wst-core on PyPI |
pip install omni-wst-core |
| Zero-copy FFI bridge (Rust ⇄ C++) | samvardhan03/Omnipulse — omni-ffi/ |
omni-ffi on crates.io |
cargo add omni-ffi |
| Rust orchestrator + indexes | samvardhan03/Omnipulse — omnipulse-rs/ |
vector-index, sliced-wasserstein on crates.io; omnipulse-mcp binary |
cargo install omnipulse-mcp |
| Python agentic control plane | samvardhan03/Omnipulse — omnipulse-agent/ |
omnipulse-agent on PyPI |
pip install omnipulse-agent |
If you want the end-to-end demo, install both Python packages and
the Rust binary in the same environment, set ANTHROPIC_API_KEY, and
run python -m omnipulse_agent.run --wav your.wav.
Try it (≈ 90 seconds)
# 1. The math engine (PyPI wheel — C++/CUDA inside)
pip install omni-wst-core
# 2. The Rust MCP orchestrator binary
cargo install omnipulse-mcp
# 3. The Python control plane (agentic layer)
pip install omnipulse-agent
# 4. Run the demo
export ANTHROPIC_API_KEY=sk-ant-...
python -m omnipulse_agent.run --wav your.wav
You will see the four-stage trace: ingest → shm pin → cxx bridge →
HNSW insert, with the 28-char SHA3 hex shm name printed at each
boundary.
Repository layout (this showcase)
.
├── README.md ← you are here
├── omni-wst-core/ ← C++/CUDA DSP engine
├── omni-ffi/ ← cxx zero-copy FFI bridge
├── omnipulse-rs/ ← Rust workspace (vector-index, sliced-wasserstein, omnipulse-mcp)
├── omnipulse-agent/ ← Python agentic control plane
└── site/ ← Next.js marketing site (omnipulse.dev)
The vendored snapshots are kept for reading only. They lag the
canonical per-module repos slightly. Always clone the per-module repo
if you intend to build or contribute.
Maintainers
- Samvardhan Singh — C++/CUDA DSP engine (
omni-wst-core), the cxx FFI bridge (omni-ffi), and the PyPI agentic control plane (omnipulse-agent). Focus: automation engineering, AI/MLOps pipelines, engineering outcomes. samvardhan.vercel.app · [email protected] - Yash Mishra — Rust crates (
vector-index,sliced-wasserstein). Focus: concurrent systems, optimal transport, real-time indexing. linkedin.com/in/mishra-yash2002 · [email protected] - Phase 3 — Autonomous Agentic Control Plane is co-authored by Samvardhan and Yash.
License
Apache 2.0 for all open source modules. A commercial license tier
exists for production deployments at scale — see
omnipulse.dev/licensing or email
[email protected].
If this repo is useful to you, star it so others can find it.
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi