AxisOS Nebula Engine — Nemotron & C99 Triad

Sovereign 3-tier agent engine uniting 112-us freestanding C99 reflexes with NVIDIA Nemotron Nano, Super & Ultra on Nebius Token Factory.

Comment

AxisOS Nebula Engine

Built by Crusan & Co. — Build a working AI system on open, independent infrastructure, and own the whole thing.

AxisOS Nebula Engine is a sovereign, 3-tier multi-agent operating engine that unites 100% freestanding, zero-libc C99 edge and cloud binaries (~112 µs reflex latency) with NVIDIA's open-source Nemotron model family (Llama-3.1-Nemotron-Ultra-253B-v1, Llama-3.3-Nemotron-Super-49B-v1, and Llama-3.1-Nemotron-Nano-8B-v1) served on Nebius Token Factory and Nebius Serverless Endpoints. Engineered under the Three Pillars of AxisOS — HONEST, AUDITABLE, and OWNED — Nebula Engine eliminates bloated middleware and stretches GPU inference credits by routing every task to the exact tier of silicon and open-source intelligence it requires.

Inspiration

Modern agent frameworks suffer from two critical flaws: they burn massive 200B+ parameter GPU calls on trivial status checks, and they trap operators inside opaque, proprietary software stacks they cannot inspect or own. When we saw the Nebius x NVIDIA Global AI Hackathon challenge — "Build a working AI system on open, independent infrastructure, and own the whole thing" — it aligned 1-to-1 with the AxisOS Sovereign Triad architecture built by Crusan & Co. We set out to prove that combining a 49 KB freestanding C99 reflex mind with NVIDIA Nemotron Nano, Super, and Ultra models on Nebius Token Factory creates a faster, leaner, and completely auditable multi-agent system.

What it does

AxisOS Nebula Engine orchestrates workloads across a 3-Tier Sovereign Routing Hierarchy:

  • Tier 0 (112 µs Freestanding C99 Reflex — zero-core :8081 & axis-agent :8082): Handles deterministic system telemetry (/proc/loadavg, /proc/meminfo, /sys/class/thermal, /proc/net/dev), RAMGAK-bracketed DDR4 memory checks, and Write-Once-Read-Many (WORM) Anti-Data-Loss vault enforcement in 112 microseconds at zero GPU token cost.
  • Tier 1 (Fast Everyday Tactical Intelligence — agent-lite :8080 & zero-comms :8085 via Nebius Token Factory): Routes high-frequency intent classification, dual-homed RP2350 mesh node heartbeats (Node 001 [💃🍪🕺]), and daily workflow synthesis to nvidia/Llama-3_1-Nemotron-Nano-8B-v1 and nvidia/Llama-3_3-Nemotron-Super-49B-v1 on Nebius Token Factory (https://api.studio.nebius.com/v1), keeping everyday calls sub-second and stretching inference credits 10x further.
  • Tier 2 (Deep Multi-Step Reasoning — axis-agent :8082 via Nebius Token Factory): Escalates complex architectural synthesis, bare-metal ARM64/x86_64 C99 syscall verification (svc #0 / syscall), and 3-NIC failover planning to nvidia/Llama-3_1-Nemotron-Ultra-253B-v1 on Nebius Token Factory.
  • Hardware-Grounded WORM Anti-Data-Loss Shield: Any attempt by a prompt or process to DELETE, OVERWRITE, or TRUNCATE the audit vault is deterministically rejected (HTTP 403 Forbidden — blocked_by_worm).

How we built it

We built AxisOS Nebula Engine by pairing our freestanding C99 Sovereign Triad binaries with an OpenAI-compatible Nebius Token Factory routing engine (src/nebula_engine/):

  • Freestanding C99 Sovereign Triad (127 KB uncompressed / 87 KB compressed FROM scratch container): Compiled with -nostdlib -ffreestanding -static using raw Linux kernel syscalls (zero-core 49 KB, lite 16 KB, axis-agent 20 KB, zero-comms 17 KB, cloud-triad-gateway 25 KB).
  • Nebius Token Factory Integration (nebius_client.py & router.py): Connects directly to https://api.studio.nebius.com/v1/chat/completions to dynamically dispatch between NVIDIA Nemotron Nano-8B, Nemotron Super-49B, and Nemotron Ultra-253B based on task complexity.
  • Live /proc & /sys Ground-Truth Bridge (triad_bridge.py): Reads real machine state (load average, RAMGAK memory brackets, SoC thermal °C, and active NIC packet counters) with zero hardcoded mock strings.
  • Automated 4-Stage Verification & Test Suite (scripts/run_nebula_demo.py & tests/test_nebula_engine.py): Verifies all three routing tiers and the WORM Anti-Data-Loss shield (6/6 unit tests passing).

Challenges we ran into

  • Balancing Sub-Millisecond C99 Reflexes with Cloud LLM Latencies: Bridging a 112 µs freestanding C99 daemon in DDR4 RAMDisk with cloud GPU inference required a strict, deterministic Tier-0 pre-filter so hardware health checks never wait on network round-trips.
  • Zero-libc Container & Edge Parity: Ensuring our 5 freestanding C99 Triad binaries ran identically on both a diskless RAM-booted Raspberry Pi 4 (pp-core-pri) and our 87 KB FROM scratch cloud outpost required clean-room syscall wrappers (svc #0 on ARM64 and syscall on x86_64).
  • Preventing Destructive Agent Hallucinations: Designing the Axis-Agent (:8082) WORM Vault at the syscall and router boundary so no LLM tool call can ever delete or overwrite historical audit logs or operator rollback checkpoints.

Accomplishments that we're proud of

  • 100% Freestanding C99 + NVIDIA Nemotron Synergy: Proving that a 49 KB C99 mind (zero-core) and a 253B open-source reasoning model (NVIDIA Nemotron Ultra on Nebius Token Factory) complement each other like a human brainstem and prefrontal cortex.
  • Extreme Efficiency: Packing all 5 Sovereign Triad cloud binaries into an 87 KB compressed FROM scratch container while cutting GPU token waste by routing routine telemetry to Tier 0 and fast everyday calls to Nemotron Nano-8B / Super-49B.
  • Hardware-Grounded Honesty: Zero fake status strings—every telemetry report is pulled live from /proc and /sys and locked into an append-only WORM ledger.

Live Sovereign Triad Outpost (Freestanding C99, zero-libc, FROM-scratch, 87 KB): https://axisos-sovereign-triad-4zye7tn3qq-uw.a.run.app

Live AxisOS Fleet Deck & Console: https://axisos.pro/console/

Public GitHub Repository (MIT Licensed): https://github.com/mcrusan/AxisOS-Nebula-Engine

What we learned

We learned that Nebius Token Factory makes tiered open-source model orchestration effortless: because NVIDIA Nemotron Nano-8B, Super-49B, and Ultra-253B share a clean, unified OpenAI-compatible API surface on Nebius, our AxisOS Sovereign Triad can switch from a lightweight 8B tactical scout to a 253B deep architect in a single parameter change while keeping full ownership of our prompts, telemetry, and infrastructure.

What's next for AxisOS Nebula Engine

  • 26 TOPS Hailo-8 Edge NPU + Raspberry Pi 5 Integration: Extending Tier-0 local vision and sensor pre-processing at the shop edge alongside Nebius Token Factory cloud escalation.
  • Nebius Serverless Jobs for Nightly Kernel & Firmware Audits: Running automated overnight static and dynamic verification sweeps across our C99 syscall tree and RP2350 firmware using NVIDIA Nemotron Ultra-253B.
  • Expanded Multi-Node Mesh: Scaling the [💃🍪🕺] UDP :50000 hardware squadron across additional autonomous field outposts.

Built With

Share this project:

Updates

Submission history