aip-engine 0.3.0

Algebraic Independence Processor - auto-detects structure for optimal computation

AIP Engine

Algebraic Independence Processor - Auto-detects structure in your data and picks the optimal computation path.

Install

pip install aip-engine

Quick start

import aip
import numpy as np

# Matrix multiply: auto-detects sparse vs dense
A = np.eye(1000) * 3.0
B = np.eye(1000) * 2.0
C = aip.multiply(A, B)  # uses scipy.sparse automatically

# Solve linear system: auto-routes by shape and density
x = aip.solve(A, np.ones(1000))

# Graph analysis: communities, fraud detection
edges = [(0, 1), (1, 2), (2, 0), (3, 4)]
info = aip.analyze(edges)

# Matrix conversion: auto-detect optimal format
report = aip.analyze_format(A)
print(report['recommended'])  # 'DIA' (diagonal detected)
sparse_A = aip.to_sparse(A)   # auto-picks best sparse format

# MoE acceleration: only activate needed experts
result = aip.moe_forward(x_input, experts, router, n_active=4)

Modules

Detector (aip.detect, aip.detect_matrix)

Analyzes matrices and graphs to detect: density, sparsity, bipartition, connected components, max degree. Routes to optimal algorithm. Complexity: O(n + m).

Matrix Operations (aip.multiply, aip.solve)

Auto-routing multiplication and linear system solving:

Input	AIP detects	Routes to
Sparse matrix	density < 30%	scipy.sparse (CSR)
Dense matrix	density >= 30%	numpy / BLAS
Square sparse system	square + sparse	spsolve
Rectangular sparse	non-square + sparse	LSQR iterative
Rectangular dense	non-square + dense	lstsq

Matrix Conversion (aip.convert, aip.to_sparse, aip.to_dense)

Auto-detects internal structure (diagonal, banded, block, symmetric, triangular) and converts to the optimal storage format:

Structure detected	Converts to
Diagonal	DIA
Banded (narrow bandwidth)	DIA
Block-diagonal	BSR
General sparse	CSR
Dense (> 30%)	numpy array

Includes aip.analyze_format() for full structural analysis with memory comparison, and aip.benchmark_convert() for actual timing benchmarks across formats.

Graph Analysis (aip.analyze, aip.detect_fraud)

Community detection, critical nodes, and fraud pattern detection (odd cycle / carousel detection) in transaction networks.

MoE Acceleration (aip.moe_forward, aip.MoEEngine)

Inference engine for Mixture-of-Experts models. Detects and activates only the needed experts (~3% of total), reducing compute and memory proportionally.

Documentation

See the docs/ folder for detailed API documentation of each module.

Requirements

• Python 3.8+
• NumPy >= 1.20
• SciPy >= 1.7

License

MIT - Carmen Esteban, 2025-2026