NEURA PARSE

00Edge runtime product

A signed software path from model to device.

NeuralOS packages Linux 6.12 LTS, PREEMPT_RT, secure OTA, model encryption, drone protocols, robotics workflows, and 12 inference backends into one deployable edge runtime.

NeuralOS v5.0.0 · Buildroot 2025.08 · NPIE v2.0.0

View on GitHub ↗Explore architecture
NeuralOS edge fleet console showing secure boot, signed OTA updates, inference health, robotics graph, and drone telemetry
Current product interface · secure boot, signed OTA, inference health, robotics graph, and autonomous-device telemetry.
Minimum RAM
64 MB
Worst-case latency
<1 ms
Boot time
<5 s
Inference backends
12
Desktop apps
33
Custom packages
46

01Operating contexts

NeuralOS is most useful when the runtime is tied to a workflow: signed releases, local inference, telemetry health, OTA rollback, and evidence that NowFlow and operators can inspect.

BVLOS drone operations console with fleet telemetry, route constraints, Remote ID, safety state, and evidence panels

Aerospace

Signed edge runtime, MAVLink/PX4 integration, Remote ID state, telemetry quality, route constraints, and post-flight evidence for drone, IHA, and UAV programmes.

Inspect the field route
Uncrewed systems assurance console with signed manifests, human authority checkpoints, runtime telemetry, and fail-safe monitoring

Defense

Model lineage, device identity, policy bundles, human authority checkpoints, and replayable exception evidence for mixed aerial and ground assets.

Inspect the field route
Robotics fleet operations console with AMR routes, digital twins, industrial robot cells, safety zones, OTA state, and telemetry

Industrial

ROS 2/DDS, OPC UA bridges, edge model packages, digital twin validation, OTA rollback, and NowFlow exception routing for AMR, AGV, cobot, and inspection fleets.

Inspect the field route

02Architecture

Six integrated layers connect the real-time kernel to networking, assurance, autonomy, inference, and operator applications.

Linux 6.12 LTS + PREEMPT_RT fully preemptible kernelBuildroot 2025.08 with 46 custom packagesNPIE v2.0.0 unified inference layer

System stack

L01–L06
L01

Applications

Neural Studio, AI Agent Hub, Drone Command Center, Quantum Lab + 29 more

L02

NPIE v2.0.0

12 inference backends -- LiteRT, ONNX, emlearn, WasmEdge, NCNN, ExecuTorch, OpenVINO, llama.cpp, whisper.cpp, stable-diffusion.cpp, MLC LLM, QuEST

L03

Robotics & Drone Stack

MAVLink 2.0, PX4 1.16.0, ArduPilot, MAVSDK, swarm engine, 16-state EKF, Remote ID

L04

Networking

WiFi mesh (802.11s + BATMAN-adv), eBPF QoS, 4G/5G cellular, WireGuard VPN, delta+LZ4 telemetry

L05

Security

Secure Boot (FIT), dm-verity, AES-256-GCM encryption, RAUC OTA with rollback, zero telemetry

L06

Kernel

Linux 6.12 LTS + PREEMPT_RT -- fully preemptible, 1000 Hz tick, SCHED_DEADLINE

03Inference layer

Neura Parse Inference Engine provides a unified API across 12 hardware-optimized backends. Deploy models trained in any framework. 5,325 inferences/sec, 4.55 GFLOPS.

01

LiteRT

02

ONNX Runtime

03

emlearn

04

WasmEdge

05

NCNN

06

ExecuTorch

07

OpenVINO

08

llama.cpp

09

whisper.cpp

10

stable-diffusion.cpp

11

MLC LLM

12

QuEST

Backend definitionsRuntime role for each of the twelve adapters
LiteRT
Google's on-device ML runtime
ONNX Runtime
Cross-platform inference
emlearn
Tiny classical ML for embedded devices
WasmEdge
WebAssembly model runtime
NCNN
High-perf mobile neural network
ExecuTorch
PyTorch edge deployment
OpenVINO
Intel hardware optimization
llama.cpp
On-device large language models
whisper.cpp
On-device speech recognition
stable-diffusion.cpp
On-device image generation
MLC LLM
TVM-compiled LLM runtime
QuEST
Quantum circuit simulation
Compute throughput
Inferences per second
Memory bandwidth

04Autonomy stack

MAVLink 2.0, PX4 Autopilot 1.16.0, swarm coordination with Raft consensus, 16-state EKF sensor fusion, and ASTM F3411 Remote ID compliance -- all pre-integrated.

Flight control, fleet coordination, and sensor fusion remain one inspectable operating path.
  1. A01

    Full MAVLink 2.0 protocol with PX4 Autopilot 1.16.0, ArduPilot, and MAVSDK for autonomous flight control and mission planning.

    PX4 Autopilot 1.16.0 integration

  2. A02

    Multi-drone coordination with formation flying, Raft leader election for fault tolerance, and distributed task allocation.

    Formation flying engine

  3. A03

    16-state Extended Kalman Filter fusing GPS, IMU, barometer, magnetometer, and optical flow for centimeter-accurate positioning.

    16-state EKF sensor fusion

Autonomy and networking inventoryProtocol, coordination, navigation, and constrained-link detail

MAVLink 2.0 & PX4 1.16.0

  • PX4 Autopilot 1.16.0 integration
  • ArduPilot support
  • MAVSDK flight control library
  • Remote ID compliance (ASTM F3411)

Swarm Intelligence

  • Formation flying engine
  • Raft leader election protocol
  • Distributed task allocation
  • Collision avoidance mesh

Sensor Fusion & Navigation

  • 16-state EKF sensor fusion
  • GPS + IMU + barometer integration
  • Optical flow positioning
  • Camera pipeline with OpenCV

Networking stack

  • WiFi mesh (802.11s + BATMAN-adv)
  • eBPF-based QoS traffic shaping
  • 4G/5G cellular connectivity
  • WireGuard VPN tunnels
  • Delta + LZ4 telemetry compression
  • MAVLink telemetry routing

05Security

Verified boot chain, encrypted models, atomic OTA updates, and zero telemetry. Every layer locked down by default.

S01

S02

S03

S04

S05

S06

Security control definitionsBoot, integrity, encryption, connectivity, updates, and privacy

Secure Boot (FIT)

Verified boot chain using Flattened Image Tree with cryptographic signature validation at every stage from bootloader to userspace.

dm-verity

Block-level integrity verification for read-only partitions. Detects any tampering with the root filesystem at runtime.

AES-256-GCM Model Encryption

AI models encrypted at rest and decrypted in secure memory. Protects proprietary inference models from extraction.

WireGuard VPN

Built-in WireGuard for encrypted device-to-cloud and device-to-device communication with minimal overhead.

RAUC OTA with Rollback

Atomic A/B over-the-air updates with automatic rollback on failure. Delta updates minimize bandwidth on constrained links.

Zero Telemetry

No data collection, no phone-home, no analytics. Complete privacy by design -- your fleet, your data, your control.

06Deployment estate

Platform targets stay visible for deployment decisions. The deeper desktop and library inventory opens only when an operator needs it.

D1Five supported platformsKVM development through Pixhawk and NVIDIA Jetson deployment
01

x86_64 (KVM)

Run NeuralOS as a KVM virtual machine for development, CI/CD testing, and cloud-edge simulation.

02

ARM64 (QEMU)

Full QEMU emulation for ARM64 development without physical hardware. Cross-compile and test on any workstation.

03

Raspberry Pi 4/5

Native support for Raspberry Pi 4 and 5 with hardware-accelerated video, GPIO access, and camera module integration.

04

Pixhawk 6X

Companion computer integration with Pixhawk 6X flight controller via MAVLink 2.0 for autonomous drone operations.

05

Jetson Orin Nano

NVIDIA Jetson Orin Nano with CUDA acceleration for high-performance edge AI inference up to 40 TOPS.

D2Operator shell and application inventoryQt5 QML shell, 33 built-in apps, and hybrid computing libraries

The lightweight desktop is purpose-built for embedded AI workloads: Neural Studio, Drone Command Center, Quantum Lab, and 30 additional applications.

Quantum computing libraries

  • QuEST -- Quantum Exact Simulation Toolkit -- full state-vector simulation
  • Qulacs -- High-performance quantum circuit simulator
  • PennyLane -- Differentiable quantum computing framework
  • Stim / QEC -- Quantum error correction simulation
  • CUDA Quantum -- NVIDIA hybrid quantum-classical programming

Application highlights

Neural Studio
AI Agent Hub
Drone Command Center
Quantum Lab
Model Manager
Swarm Controller
Sensor Dashboard
Flight Planner
Terminal
File Manager
System Monitor
Network Manager
WireGuard VPN
OTA Updater
Camera Viewer
Log Viewer
+ 17 more applications

08Build and release

Choose the profile that matches your deployment. Each includes only the packages required for its use case.

The release record travels with the image instead of ending at compilation.
  1. R01

    Choose the hardware platform and the smallest build profile that fits the mission.

    Input · platform + profile

  2. R02

    Buildroot resolves the kernel, runtime, packages, and target-specific image.

    Artifact · reproducible image

  3. R03

    Release identity, boot policy, model package, and known limits are checked before promotion.

    Gate · release record

  4. R04

    Flash locally or move the accepted package through the controlled OTA path.

    Output · device release

  5. R05

    Runtime health, telemetry, incidents, and rollback state remain available to operators.

    Evidence · health + rollback

01

~128 MB

Custom embedded appliances

02

~256 MB

Autonomous drone platforms

03

~320 MB

Robotic and edge AI devices

04

~512 MB

Development, testing, full-featured deployments

Profile package definitionsWhat each image includes and excludes
minimal
Base system with kernel, init, networking, and security. No AI or drone stack.
drone
Minimal + MAVLink, PX4, swarm engine, sensor fusion, camera pipeline, Remote ID.
robot
Minimal + NPIE inference engine, OpenCV, llama.cpp, sensor fusion, navigation stack.
full
Everything: NPIE, drone stack, quantum libs, Qt5 desktop shell with 33 applications.

Quick start

Build console
# Clone the repository
$ git clone https://github.com/neuraparse/neuraos.git
$ cd neuraos
# Build for your target platform & profile
$ make neuraos_rpi5_drone_defconfig
$ make -j$(nproc)
# Flash and boot in <5 seconds
$ dd if=output/images/sdcard.img of=/dev/sdX bs=4M

Deploy edge AI

Pick a build profile, flash the target, and preserve the runtime and rollback evidence around Linux 6.12 LTS and the NPIE inference layer.

View on GitHub ↗Read documentation