vulgaris 0.1.3

Foundational industrial AI model for real-time telemetry, IoT, and time-series

VULGARIS

A foundational model for streaming industrial intelligence.

---

VULGARIS is an open-source foundational model designed for real-time telemetry, sensor fusion, and time-series intelligence across industrial environments. It is written entirely in NumPy with a custom reverse-mode automatic differentiation engine — no PyTorch, no TensorFlow, no CUDA dependency.

The architecture combines state-space recurrence, causal graph learning, hierarchical memory, and domain-adaptive hypernetworks into a single unified model that trains on a CPU and deploys on edge hardware.

Architecture

Module	Role
ASE — Adaptive Signal Embedding	Morlet wavelet filterbank; converts raw channels to latent sequences
HTD — Hierarchical Timescale Decomposition	Four-level parallel SSM; captures dynamics from milliseconds to minutes
SSSR — Selective State-Space Recurrence	Diagonal ZOH state-space core; selective gating per timestep
DAH — Domain-Adaptive Hypernetwork	LoRA-style adapter generation; zero-shot domain switching at inference
CRG — Causal Routing Graph	NOTEARS-constrained DAG; learns causal structure from data online
HMB — Hierarchical Memory Bank	VAE-compressed episodic slots; retrieval-augmented state transitions
ESE — Explainability Engine	CART rule extraction from latent activations; human-readable decisions
Safety	Control Barrier Function head; certified safe action projection
SHCAL	EWC + Hebbian continual learning; prevents catastrophic forgetting
CMLA	InfoNCE cross-modal latent alignment; multi-sensor fusion

Installation

pip install vulgaris

Optional extras:

pip install "vulgaris[serve]"   # REST inference server (FastAPI + uvicorn)
pip install "vulgaris[train]"   # Training utilities (tqdm, rich)
pip install "vulgaris[all]"     # Everything

Requirements: Python ≥ 3.10, NumPy ≥ 1.26, SciPy ≥ 1.11, PyYAML ≥ 6.0

Quick Start

import numpy as np
from vulgaris import Vulgaris, ModelConfig, Tensor

config = ModelConfig(
    input_dim=9,    # input channels
    n_classes=5,    # classification targets (set 0 for regression)
    output_dim=1,
)

model = Vulgaris(config)

# Batch inference: (batch, channels, timesteps)
x = Tensor(np.random.randn(8, 9, 64).astype(np.float32))
output, aux = model(x)
print(output.data.shape)  # (8, 5)

Streaming Inference

VULGARIS processes one timestep at a time with stateful recurrence — suitable for hard real-time systems:

state = model.init_state(batch_size=1)

for t in range(sequence_length):
    x_t = Tensor(sensor_reading[t].astype(np.float32))  # (1, 9)
    output_t, state = model.step(x_t, state)

Training

from vulgaris import (
    Vulgaris, ModelConfig,
    TrainingPipeline, VulgarisLoss,
    SpectralAdamW, CosineSchedule,
)

config    = ModelConfig(input_dim=9, n_classes=5)
model     = Vulgaris(config)
loss_fn   = VulgarisLoss(config)
optimizer = SpectralAdamW(model.parameters(), lr=3e-4)
scheduler = CosineSchedule(optimizer, warmup_steps=1000,
                           max_steps=50_000, min_lr=1e-6)
pipeline  = TrainingPipeline(model, config, loss_fn, optimizer, scheduler)

metrics = pipeline.train_step(x_np, y_np)

Save and Load

model.save("checkpoints/my_run")
model = Vulgaris.load("checkpoints/my_run")

Checkpoints store weights.npz, config.yaml, and metadata.json — no pickle, no binary blobs.

Configuration

All hyperparameters are dataclasses — composable and serialisable to YAML:

from vulgaris import ModelConfig
from vulgaris.config import SSSRConfig, DAHConfig

config = ModelConfig(
    input_dim=16,
    n_classes=10,
    sssr=SSSRConfig(state_dim=512, n_heads=16),
    dah=DAHConfig(n_domains=64, adapter_rank=32),
)
config.to_yaml("config.yaml")

Load from environment variables for containerised deployments:

# VULGARIS_INPUT_DIM=16 VULGARIS_N_CLASSES=10 VULGARIS_D_MODEL=256
config = ModelConfig.from_env()

Inference Server

export VULGARIS_INPUT_DIM=9
export VULGARIS_N_CLASSES=5
vulgaris-serve

Endpoints: POST /predict · GET /health · GET /metrics · GET /versions

Supports API key authentication (VULGARIS_API_KEYS), Prometheus metrics exposition, and three-level graceful degradation.

Docker

docker compose up

Brings up the inference server and a Prometheus sidecar. Memory limit: 2 GB.

Reproducibility

from vulgaris import set_seed
set_seed(42)

Self-Supervised Pretraining

from vulgaris.training import SelfSupervisedTrainer

trainer = SelfSupervisedTrainer(model, in_channels=9, d_model=64)
metrics = trainer.pretrain_step(x_np)  # masked reconstruction + temporal InfoNCE

Drift Detection

from vulgaris import DriftDetector

detector = DriftDetector(window_size=200)
detector.set_reference(reference_data)
result = detector.update(new_batch)

if result["drift_detected"]:
    print(f"KS={result['ks_stat']:.3f}  MMD={result['mmd_stat']:.3f}")

Multi-Domain Adaptation

# Switch domain at inference time — no retraining
output, aux = model(x, domain_idx=3)

Repository Structure

vulgaris/
├── vulgaris/       public API and config
├── engine/         autograd engine: Tensor, Module, layers
├── modules/        ASE · HTD · SSSR · DAH · CRG · HMB · ESE · Safety · CMLA · SHCAL
├── model/          full Vulgaris model assembly
├── training/       loss · optimizer · pipeline · conformal · self-supervised
├── inference/      FastAPI server · streaming
├── serve/          auth · metrics · degradation · versioning
├── monitoring/     drift detection
├── federated/      federated continual learning with differential privacy
├── benchmarks/     baselines · ETT · NAB · EdgeTelemetry dataset loaders
└── tests/          unit tests with numerical gradient verification

Documentation

Full technical documentation: doc.md

Citation

@software{vulgaris2026,
  title   = {VULGARIS: Foundational Model for Streaming Industrial Intelligence},
  author  = {Khan, Jalaluddin},
  year    = {2026},
  url     = {https://github.com/jalaluddinkhan1/vulgaris},
  license = {Apache-2.0},
}

License

Apache License 2.0. See LICENSE for details.