GSaaS.ai // Ground Station as a Service

01 THE ENGINE ROOM

Meet DAVE — the brain in the bunker.

DAVE is not an API key. DAVE is a physical machine. A multi-GPU inference cluster with 512 GB of ECC RAM, fed by fiber, cooled by overkill, and stitched together by one engineer who got tired of renting other people’s compute. Every model, every byte of spectrum, every decision — stays on premises.

RACK · DAVE-01 INFERENCE NODE TEMP 41°C

U24

GPU// CUDA · 4× tensor accelerators

71% · 511 W

U22

GPU// LLM serving · Dave Chat Bots active = 2

KV Cache: 42.8 / 80.0 GB

U18

CPU// Threadripper 32 cores

52% · 145 W

U16

MEM// 512 GB DDR5 ECC

318 / 512 GB

U12

NVMe// Captures/Anomalies

0.532 TB used · 12 anomalies

U08

SDR I/O// Ettus USRP

2 active streams · 980 MB/s

U04

UPS// 7 kVA · grid-isolated

line OK · runtime 3h 20m

// philosophy.txt

The cloud is somebody else’s computer. Spectrum is a physical thing — it arrives at an antenna at a real address, in real time, with real latency budgets. So the brain that processes it lives in the same room as the radio. No round-trips. No telemetry leaking out. No per-token bill.

DAVE runs open-weight LLMs, vision models, and custom DSP pipelines side by side. Specialised AI agents live on top — each one a small program that knows one job: demodulating a downlink, classifying a transient, tracking a bird, or just sanity-checking the other agents.

LOCAL-FIRST

Every model weight, every spectrum capture, every inference — on metal I own.

REAL-TIME

Sub-second turnaround from antenna to agent decision. No cold-start cloud.

SOVEREIGN COMPUTE

No cloud dependencies. No third-party APIs. No per-token billing. True operational independence.

RAW DATA

I/Q samples saved end-to-end. Reprocess yesterday’s sky with tomorrow’s model.

// HW_01 : DAVE (CURRENT STATE, JUST GOT OUT OF BED LOOK)

// HOVER FOR REAL HW

// HW_01v2 :: DAVE (IF HE EVERS GETS A PROPER JOB)

[ACCELERATION] :: Up to 192GB VRAM Quad discrete workstation-class tensor processors for hosting massive, unquantized open-weight models entirely in memory.

[COMPUTE] :: 51,200 CUDA & 1,600 Tensor Cores Massive parallel processing grid distributing complex signal analysis and inference loads across tens of thousands of execution units.

[SOVEREIGNTY] :: 100% On-Prem, Offline, Non-Cloud Zero reliance on external APIs. No telemetry leaving your network. Fully functional in completely air-gapped environments.

[LIFECYCLE_SYNC] :: Evolution Continuity Silicon architectures evolve relentlessly. DAVE is engineered to ensure your intelligence layer perfectly mirrors the lifecycle of the physical equipment it operates. No forced cloud deprecation schedules—just a sovereign control plane that evolves precisely when your hardware does.

// HW02 :: GSaaS-AI-Edge (PROTOTYPED)

The GSaaS AI Edge is a standalone, high-performance ARM network and compute appliance designed to bridge the physical gap between your SDRs and your local infrastructure. Built around a hardware Packet Processing Engine and a 45 TOPS Neural Processing Unit, it provides massive bandwidth for routing I/Q streams and raw, localized compute for your own ML pipelines. No cloud APIs, no subscriptions, no telemetry leaving your network. It simply buffers, processes, and routes your satellite downlinks exactly where they land.

// HOVER FOR REAL HW

[RF_STACK] :: 2x RX & 2x Signal Source Generators Integrated Software Defined RF stack featuring dual independent receive channels and dual signal source generators for localized spectrum synthesis.

[MEMORY] :: Up to 48GB DRAM Ring Buffer High-capacity volatile memory pool for continuous, zero-latency I/Q data buffering.

[STORAGE] :: Up to 16TB Permanent Storage High-speed NVMe solid-state storage for localized data retention and continuous spectrum recording.

[INFERENCE] :: 45 TOPS AI Accelerator Embedded Neural Processing Unit (NPU) providing raw compute overhead for deploying your own local machine learning models.

[OS_CORE] :: Native ARM OS Clean, bare-metal operating environment optimized for hardware interrupts and long-term stability.

[INTERFACE] :: 1920x480 60Hz Ultra-Wide Touchscreen Integrated tactile display for real-time localized alerting and system monitoring.

[NET_INTERFACES] :: 2x 10GbE SFP+ & Quad 2.5GbE Primary network ingress and high-bandwidth local routing matrix for external SDRs and telemetry kit.

[EMBEDDED_INTERFACES] :: Wi-Fi 7 & Global 5G/LTE Embedded wireless and cellular modems providing automated failover and out-of-band management.

[ROUTING] :: Integrated SD-WAN Intelligent SD-WAN routing across all available interfaces to ensure reliable data delivery to the DAVE cluster.

[OFFLOAD] :: Hardware Packet Engine (PPE) Dedicated silicon for packet routing, NAT offload, and direct SFP daisy-chaining, freeing the CPU entirely for your workloads.

[POWER] :: Detachable 100Wh Battery Pack Hot-swappable power cell explicitly sized to standard approved Airline compliant 100Wh battery capacity.

02 THE ARSENAL · HARDWARE & TOOLS

Antennas, radios, rotors — and the code that drives them.

The ground station is a deliberately weird mix of consumer SDRs, lab-grade RF kit, and a lot of duct-taped Python. Everything is scripted, tracked, and logged. Nothing is “managed” by someone else.

[ MORE HERE ]

DIGITAL TWIN

The Virtual Antenna Installation

How we create a digital installation for the AI to thrive in.

3D modelling
Photogrammetry
Lidar survey
Ephemeris Engines

[ MORE HERE ]

TRACKING

Autonomous Pass Pipeline

TLE ingestion to motor commands, fully unattended.

SGP4 propagation · live TLEs
Az/El rotor · sub-degree
Doppler pre-comp on RX
Scheduler queues 24h of passes

[ MORE HERE ]

COMPUTE

DAVE Inference Cluster

The bunker brain. Where IQ samples become decisions.

Multi-GPU · CUDA + TensorRT
512 GB ECC DDR5
32 TB NVMe spectrum buffer
10 GbE fabric, no cloud egress

[ MORE HERE ]

DSP

Signal Processing Stack

GNU Radio flowgraphs, custom C++ blocks, and a lot of FFT.

Live waterfall + classifier
Demods: FSK · GMSK · BPSK · CCSDS
Noise-floor + RFI tagging
End-to-end IQ archival

[ MORE HERE ]

Local Model Zoo

Open-weight LLMs and vision/audio models hosted on DAVE.

LLM serving · vLLM / llama.cpp
Spectrogram CNN classifiers
Whisper-class audio decode
RAG over capture archive

[ MORE HERE ]

OPS

Control Plane

Everything observable, scriptable, and version-controlled.

Prometheus + Grafana dashboards
Ansible-managed nodes
Git-tracked flowgraphs
Self-hosted notebooks

[ MORE HERE ]

INTERNAL_API // OVERMIND

Cluster Command

Master orchestration logic, live telemetry, and global agent tasking. The headless control plane.

GSaaS Command Center
Master Router (/docs/agents)
DAVE Monitor (/api/stats)
Heavy Lifter Agent

[ MORE HERE ]

INTERNAL_API // FLEET

Agent Swarms

Isolated autonomous workforce logic. Pools dedicated to real-time classification and decryption.

SIGINT Analyst Fleet
Cryptanalysis Fleet
Task Force Alpha (LLM)
Task Force Beta (Logic)

[ MORE HERE ]

INTERNAL_API // RAW

Signal Pipeline

Direct high-bandwidth taps into the hardware layer. Access unfiltered spectrum buffers.

Core Gateway Router
Raw DSP Node Control
Live TCP SDR Ingest Stream
Full Spectrum NVMe Dump

[ MORE HERE ]

INTERNAL_API // DATA

Deep Archive

Programmatic query interfaces for historical signal captures and system endpoints.

General System API Docs
Historical Capture Archive

// agent_fleet · running on DAVE --:--:-- UTC

spectrum_watcher Continuous wideband sweep · flags anomalies vs baseline. 74% RUNNING

pass_scheduler Ingests TLEs, predicts passes, commands the rotor. 22% RUNNING

downlink_decoder Demod + frame sync · CCSDS, AX.25, custom protocols. 88% BUSY

doppler_tracker Real-time frequency correction from orbital model. 18% RUNNING

classifier_cnn Spectrogram CNN · labels modulation + emitter class. 55% RUNNING

rag_librarian Indexes captures + notes · answers operator queries. 6% IDLE

watchdog Cross-checks the other agents. Trusts nothing. 9% RUNNING

Catching signals
from space.
Feeding them to DAVE

⚠️ DON'T LOOK UP! in the skies above you your location right now…