Hugging Face's logo

rootcastleengineering
/
sofia

English
sofia-engine
edge-ai
industrial-ai
scientific-computing
embedded-ai
signal-processing
digital-signal-processing
predictive-maintenance
condition-monitoring
vibration-analysis
anomaly-detection
industrial-iot
iiot
telemetry
edge-computing
tinyml
on-device-learning
embedded-systems
machine-health
time-series
python
typescript
c
Model card Files
xet
 Community
sofia / README.md
rootcastleengineering's picture
rootcastleengineering
docs: add Hugging Face storage bucket reference
adf4304
preview code
|
Raw
History Blame Contribute Delete
11 kB
metadata 
license: apache-2.0
doi: 10.57967/hf/10549
language:
  - en
tags:
  - edge-ai
  - industrial-ai
  - scientific-computing
  - embedded-ai
  - signal-processing
  - digital-signal-processing
  - predictive-maintenance
  - condition-monitoring
  - vibration-analysis
  - anomaly-detection
  - industrial-iot
  - iiot
  - telemetry
  - edge-computing
  - tinyml
  - on-device-learning
  - embedded-systems
  - machine-health
  - time-series
  - python
  - typescript
  - c
library_name: sofia-engine

Sofia Engine — Scientific & Edge Intelligence Runtime

Developed by Rootcastle Engineering & Innovation
Canonical Source Repository: github.com/rootcastleco/sofia-ai
Project Homepage: rootcastle.com
DOI: 10.57967/hf/10549
Storage Bucket: huggingface.co/buckets/rootcastleengineering/sofia-bucket

1. Overview

Sofia Engine is an open-source scientific and edge intelligence runtime developed by Rootcastle Engineering & Innovation for deterministic physical telemetry processing, industrial signal analysis, machine-health diagnostics, bounded edge inference, and safety-gated technical automation.

Operating under the engineering principle "Evidence beats claims", Sofia Engine combines:

  • Scientific DSP: FFT, Welch PSD, Hilbert analytic amplitude envelope, and Fortescue 3-phase symmetrical components.
  • Physical Telemetry Ingestion: Bounded, unit-bearing telemetry frames with quality status (GOOD, DEGRADED, SATURATED).
  • Versioned Feature Extraction: 14+ statistical and spectral features packaged in ordered, reproducible FeatureVector contracts (Schema v3.0).
  • Evidence-Based Diagnostics: Noisy-OR evidence aggregation and machine health scoring with explicit uncertainty bands ($0-100 \pm \Delta$).
  • Embedded & Edge Execution: Hardened virtual machine (SofiaAsmVM), C99 zero-allocation microcontroller implementation, and experimental in-situ backpropagation.
  • Safety & Policy Gating: Strict default-DENY policy engine, Nonce + TTL replay protection, and an actuation firewall that strictly isolates external LLMs from physical machinery.

Canonical Pretrained Checkpoint Status: No canonical pretrained Sofia checkpoint is distributed in this release. This repository provides the Sofia model/runtime specification, machine-readable manifests, reproducible examples, and future checkpoint distribution interfaces. When pre-trained foundation models are released, they will be published with complete training datasets, evaluation reports, and SHA-256 digests.

2. Subsystem Maturity Table

Maturity levels are assigned strictly based on verifiable implementation and test coverage:

Subsystem Status Verification & Evidence
Scientific DSP Stable Parseval energy conservation verified; exact amplitude recovery to $< 10^{-14}$; 5 golden vector tests passed.
Telemetry Runtime Stable Memory capacity ceilings ($\le 65536$), timestamp plausibility, unit alias normalization, IEEE 754 non-finite rejection.
Diagnostic Engine Stable Deterministic Noisy-OR fusion; dynamic uncertainty intervals; quality attenuation on degraded signals.
Policy Engine & Replay Guard Stable Default-DENY posture; Nonce + TTL monotonic replay protection; LLM actuation firewall; secret scrubbing.
Embedded C99 Runtime Stable Zero dynamic heap allocation post-init (malloc prohibited); Q16.16 fixed-point arithmetic; static ceilings.
Sofia Assembly VM Beta Memory and register bounds checking; cycle ceilings; structured execution result (VMExecutionResult).
In-Situ Neural Training Experimental 2-layer analytical backpropagation in virtual assembly; validated against finite differences ($< 10^{-5}$ error).
Cross-Language Conformance Beta Python, TypeScript SDK (@rootcastle/sofia-engine), and C99 verified against identical golden vectors.
LLM Copilot Integration Optional Strictly advisory technical assistance; isolated from deterministic control core and physical actuators.
Quantum Emulation Experimental Educational statevector simulation and VQE parameter exploration.
Native Code Generation Experimental Preliminary x86_64 AVX2, ARM Cortex-M Thumb-2, and WebAssembly emission stubs.

3. Architecture & Data Flow 

flowchart LR
    S["Physical Sensors"] --> I["Telemetry Ingestion"]
    I --> Q["Validation & Signal Quality"]
    Q --> B["Bounded Buffers"]
    B --> D["Scientific DSP"]
    D --> F["Versioned Feature Vector"]
    F --> M["Inference Backends"]
    M --> E["Evidence Fusion"]
    E --> H["Diagnostics & Health"]
    H --> P["Policy Engine"]
    P --> O["Advisory / Controlled Output"]

Key Architectural Invariant: External LLMs (OpenAI, NVIDIA NIM, OpenRouter) operate exclusively as optional copilots at Layer 10. They have no read or write access to the deterministic signal processing or policy enforcement layers.

4. Scientific Scope & Calculations

Sofia Engine implements calculations relevant to industrial engineering standards:

  • Mechanical Vibration: Implements calculations relevant to ISO 10816 / ISO 20816 vibration severity evaluation (RMS, peak, crest factor, kurtosis, skewness).
  • Spectral Analysis: Discrete Fourier Transform (one-sided FFT magnitude), Welch's averaged periodogram with window gain corrections, spectral centroid, spectral entropy, and spectral flatness.
  • Envelope Demodulation: Hilbert transform analytic signal for bearing defect and gear-mesh modulation extraction.
  • Electrical Power Quality: Implements calculations relevant to IEEE 519 and IEC 61000-4-30 analysis, including True RMS, Active/Reactive/Apparent Power, Total Harmonic Distortion (THD), and Fortescue 3-phase symmetrical components ($V_0, V_1, V_2, \text{VUF}$).
  • Process Telemetry: Thermal rates of change ($dT/dt$), pulsation peak-to-peak, and fluid pressure crest factors.

Note: Software implementation does not constitute formal laboratory certification. Operational deployment requires qualified engineering calibration.

5. Installation

Core Runtime (NumPy Only) 

pip install sofia-engine

Industrial Telemetry Integrations (MQTT, Modbus, Serial) 

pip install "sofia-engine[industrial]"

From Canonical Source 

git clone https://github.com/rootcastleco/sofia-ai.git
cd sofia-ai
pip install -e .

6. Reproducible Examples

Executable scripts are maintained in the examples/ directory:

  • examples/basic_inference.py: Signal window packaging, 14-feature extraction, threshold anomaly detection, and evidence-based health scoring.
  • examples/signal_analysis.py: FFT magnitude, Welch PSD with Parseval energy conservation check, Hilbert analytic envelope, and Fortescue symmetrical components.
  • examples/edge_runtime.py: Bounded SignalFrame, Sofia Assembly VM execution, in-situ neural training step, and safe sofia.model.v1 serialization.

To execute the basic inference example:

python examples/basic_inference.py

7. Model Manifests & Checkpoint Policy

Hugging Face distribution files are machine-readable and schema-validated:

Checkpoint Policy

  • Zero-Pickle Invariant: Model weights must be distributed in .safetensors or .npz format. Python pickle is strictly prohibited.
  • Cryptographic Verification: Every model artifact must declare its lowercase SHA-256 parameter digest.
  • See artifacts/README.md for full details.

8. Empirical Performance Benchmarks

Measured on Profile A hardware (Intel64 x86_64, Windows 11, Python 3.12.10 CPython) across $N = 100$ iterations:

Metric Measured $p_{50}$ Target Budget Result
FFT / Welch PSD (1024 samples) $307.6,\mu\text{s}$ $\le 500,\mu\text{s}$ PASS
FFT / Welch PSD (4096 samples) $387.4,\mu\text{s}$ $\le 2000,\mu\text{s}$ PASS
Analytic Envelope (2048 samples) $127.8,\mu\text{s}$ $\le 1000,\mu\text{s}$ PASS
Symmetrical Components (Fortescue) $55.5,\mu\text{s}$ $\le 1500,\mu\text{s}$ PASS
Peak Memory Consumption $7.14\text{ MB}$ $\le 32\text{ MB}$ PASS

Complete benchmark reports and JSON schema are available in benchmarks/.

9. Safety & Limitations

  • Engineering Judgment: Sofia Engine outputs are diagnostic aids. They do not replace physical inspection or qualified engineering sign-off.
  • Physical Safety Interlocks: Under no circumstance should AI or diagnostic outputs directly control machinery without external, hardware-certified safety interlocks (e.g. SIL relays).
  • Sensor Calibration: Diagnostic accuracy is bounded by sensor fidelity, mounting configuration, and signal acquisition quality.
  • Read docs/LIMITATIONS.md and docs/SECURITY.md for detailed boundaries.

10. Citation

To cite Sofia Engine in scientific software, publications, or engineering documentation:

@software{sofia_engine_2026,
  author       = {{Rootcastle Engineering \& Innovation}},
  title        = {Sofia Engine: Scientific \& Edge Intelligence Runtime},
  year         = {2026},
  version      = {3.0.0a1},
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/10549},
  url          = {https://huggingface.co/rootcastleengineering/sofia}
}