ψ Vikshep

Scientific compute infrastructure

Deterministic features for physics. Nothing learned, nothing leaked.

pip install vikshep
Talk to founders →

AGPL-3.0 + Commercial · open source

WAVELET SCATTERING · |x ⋆ ψλ|

A signal, decomposed across scales.

Fixed filters, no training. S₁[λ] = |x ⋆ ψλ| ⋆ φJ — each row is one scale band. Bright cells are where the signal has energy. Nothing is learned.

Why every modern tagger quietly breaks discovery

A neural net trained on jet features learns the jet mass — and your bump-hunt with it.

ATLAS / CMS

Mass decorrelation is a top concern in every boosted-object analysis

HL-LHC 2030

Restart with ×5 luminosity, ×10 pile-up, an order of magnitude more data

BSM searches

Anomaly detection is the field's response to not knowing what to look for

Two ideas, one guarantee

Deterministic features. Closed-form decorrelation. Reusable across every scientific domain.

ψ

Deterministic feature extraction

Fixed analytic Morlet wavelets cascade through the data — translate, modulate, pool. No learned weights means no mass leakage by construction. r₂ = S₂/S₁ is dimensionless and scale-invariant.

xx
xψλ1|x \star \psi_{\lambda_1}|
S1S_1
j0j1j2j3
ψλ2|\cdot \star \psi_{\lambda_2}|
S2S_2

Mathematical decorrelation

A weighted distance-correlation penalty enforces statistical independence between the tagger output and the resonance mass. dCorr = 0 is a closed-form guarantee — not a heuristic.

1000150020002500m_jj [GeV]0.20.40.60.8backgroundsignal

r₂ = S₂/S₁. r₂ is dimensionless — it can't carry an energy scale. It's the scattering analogue of D2, generalized to a full basis.

One engine, every dimension

SE(2) for 2-D jet images. SO(3) for 3-D density fields. 1-D Morlets for GW strain. The (Dim, Group, J, Q, L) parameters are runtime config — not a rebuild.

The filter bank

J scales, L orientations — fixed by design.

Each cell shows the energy of one Morlet wavelet in (scale j, orientation θ) space. Toggle J and L to explore coverage — no recompilation needed.

J4
L8

4×8 filter bank · 32 wavelets

The bank that the engine uses. Fixed analytic Morlets at log-spaced scales and uniformly-spaced orientations. Filters never change — that's the point.

The mass-sculpting problem

NN taggers sculpt. r₂ doesn't.

Every standard jet classifier correlates with m_jj — cutting on it deforms the background mass spectrum. The scattering ratio r₂ = S₂/S₁ is mass-decorrelated by construction. DisCo makes it rigorous.

50010001500200025003000m_jj [GeV]0.0000.0200.0400.0600.080pre-cut bkgpost-cut bkg
cut c0.50
λ DisCo0.50

0.958

AUC

signal vs bkg

2.69

JSD ×10³

mass sculpting

The standard NN tagger is correlated with jet mass — cutting on it sculpts the background spectrum (JSD rises). The Vikshep r₂ tagger with DisCo penalty (λ) keeps the post-cut shape flat. AUC quantifies discrimination power; JSD quantifies sculpting. Drag the sliders to see the tradeoff.

Three recipes, one engine

Drop in a pipeline. Or write your own.

01

HEP Tagging (DisCo)

Mass-decorrelated boosted-object tagging.

Ingest jet constituents or calorimeter images, extract r₂ substructure features via 2-D oriented scattering, train the classifier under a distance-correlation penalty on the resonance mass. Background shape preserved; tagging efficiency competitive with the best learned taggers.

bun run main.ts process --request "tag jets, decorrelate mass"
Click to know more →
02

BSM Anomaly Detection

Template-free new-physics search.

Embed every event in the scattering feature space; index the SM background in an HNSW graph with Sliced-Wasserstein distance; flag events beyond a calibrated distance from the SM manifold. No signal model, no retraining.

τ = 2.5
bun run main.ts process --request "find events that don't look like SM"
Click to know more →
03

General Feature Extraction

Cosmology, plasma, GW, hydrodynamics.

Pass any {dim, group, J, Q, L} config at runtime. 1-D for transient time series. 2-D for lensing/CMB. 3-D for density fields and plasma simulations. Same engine, same MCP tool, different physics.

bun run main.ts process --request "extract rotation-invariant features"
Click to know more →
04Preview

Fast-Sim Validation

Preview · Q3 2026

A principled fidelity metric for ML-generated Geant4 surrogates — Wasserstein distance between scattering distributions of generated vs full-simulation showers. Decomposable by scale and orientation.

Coming Q3 — Talk to founders →
Click to know more →

How it works

The full stack, from TypeScript to CUDA.

ψ TypeScript / Bun agent (Vikshep)
Coordinator · MCP client · Recipe library
React preview dashboard
28-hex shm OID · JSON-RPC 2.0 over stdio
🦀 omnipulse-mcp (Rust — Data Plane)
shm_open + mmap · spawn_blocking
HNSW + Sliced-Wasserstein
u64 host pointer · cxx zero-marshalling
🌉 omni-ffi (Rust ⇄ C++ bridge)
cxx 1.0 — unsafe extern "C++"
CPU and CUDA dispatch paths
pinned host page
⚙️ omni-wst-core (C++/CUDA)
ScatteringEngine<Arch, Dim, Group, J, Q, L>
Morlet bank · Solid-harmonic bank

Every arrow is one of: a 28-char hex OID, a \n-terminated JSON frame, or a registered host page.

Where this earns its license

Built for analyses where the statistical guarantee is the publishable result.

Particle physics

Mass-decorrelated jet tagging for ATLAS, CMS, and HL-LHC-era analyses. Validated against your own Wilks Δχ² significance pipeline.

Wilks Δχ² · JSD-preserved background · zero data egress

Cosmology & astrophysics

Rotation-invariant features for weak-lensing maps, CMB patches, and 3-D density fields. Second-order coefficients capture non-Gaussian structure the power spectrum misses.

SE(2) and SO(3) invariance · GPU-accelerated

Simulation science

Solid-harmonic scattering for SO(3)-covariant features of plasma, nuclear, and hydrodynamics fields. Runtime (Dim, Group) config — no rebuild per domain.

ROOT, HDF5, VTK loaders · no model retraining

Custom deployment · Contract-based · SLA committed

Talk to founders →

No black box

Every coefficient is derivable from the input. Every guarantee is provable.

Read the full math →

Currently in the wild

University of Edinburgh — boosted di-boson resonance search

In partnership with a nuclear-physics researcher at the University of Edinburgh, Vikshep is running on Geant4-simulated ATLAS data, replacing a neural-network tagger trained on eight high-level kinematic variables. Two numbers will be reported: the significance gain (Δσ) over the NN baseline, and the reduction in background mass-shape distortion (ΔJSD). Results: Q1 2026.

Pilot details →

Get started in three steps

Try it in 90 seconds.

# 1. Install the engine (binary wheel)
pip install vikshep

# 2. Install the Rust orchestrator
cargo install omnipulse-mcp

# 3. Run a recipe
export OMNIPULSE_MCP_BIN=$(which omnipulse-mcp)
bun run vikshep/agent/src/main.ts process \
  --input data/jets.root \
  --recipe hep-tagging-disco
View the GitHub repo →

Free for research and open-source use under AGPL-3.0. Commercial or proprietary deployment requires a commercial license — talk to founders.

Run it yourself

AGPL-3.0. Open-source. GPU-accelerated binary wheels on PyPI. Works on your hardware, in your facility.

GitHub →

Or have us run it with you

We integrate into your existing analysis pipeline, validate against your significance benchmark, and deliver the two numbers that matter.

Talk to founders →

Founders & creators

The people behind the plane

ARCHITECT — APPLIED AI / MLOPS

Samvardhan Singh

Automation Engineering & AI/MLOps Research, NielsenIQ

Automation engineering, AI/MLOps pipelines, engineering outcomes.

ARCHITECT — SYSTEMS / OPTIMAL TRANSPORT

Yash Mishra

Senior Software Engineer, Bajaj Finserv

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

RESEARCH PARTNER — ATLAS / DIBOSON PILOT

Komal Papanwar

MS Nuclear & Particle Physics, University of Edinburgh

Leads the ATLAS boosted di-boson resonance pilot analysis on Geant4-simulated data.