Open Source Ecosystem

All core libraries extracted from this project are published as standalone open-source packages:

Rust — crates.io

Crate	Description
	Neuromodulator dynamics (dopamine, cortisol, acetylcholine, tempo)
	Homeostatic reward computation for cyber-physical systems
	Sensor → spike train encoding (Poisson, temporal, predictive)
	Pluggable SNN backend trait (Rust / ZMQ IPC)
	Q8.8 parameter export + UART spike readback for FPGA
	SNN-based sparse domain routing (Anti-Hallucination Layer)

Julia — JuliaHub / General Registry

Package	Description
SpikenautLSM.jl	GPU-accelerated sparse Liquid State Machine (cuSPARSE + OU-SDE)
SpikenautNero.jl	Multi-lobe relevance scoring with cross-inhibition
SpikenautDistill.jl	Monte Carlo SNN training + FPGA distillation pipeline
SpikenautSignals.jl	Streaming Hurst / Hawkes / GBM-surprise feature extraction

SystemVerilog — GitHub

Repo	Description
spikenaut-core-sv	Parameterized Q8.8 LIF + STDP IP cores
spikenaut-bridge-sv	UART neural-cortex protocol IP
spikenaut-soc-sv	Complete reference SNN SoC for Basys3 / Artix-7

🦁 Spikenaut-SNN-v2

The Lion That Survives

Spikenaut was born in January 2026 — completely by accident.

I started university thinking I would go to medical school or even law.
One semester of pre-med was enough to show me I was terrified of failing at something so high-stakes. I felt I wasn’t smart enough, wasn’t cut out for it.
So I switched to business — hoping it would be safer.
But I quickly saw I’d just be another Business major lost in a sea of MBAs. I didn’t want to disappear.
I moved to computer science — excited about building things, coding, creating.
Then the AI hype wave hit hard. Everyone said “AI is going to replace all the coding jobs.” I believed it. I panicked.
I feared I’d spend years learning something that would vanish before I could even start.
That fear pushed me again — this time to electrical engineering. If software was going to be automated, maybe hardware was the last place left where I could build something real, something physical, something that couldn’t be replaced overnight.

But the transfer was brutal.
Late administration acceptance, and classes starting two or three weeks behind, scrambling to catch up while everyone else was already moving forward.
I struggled terribly.

Through all those pivots, discouragements, and fears, one thing stayed constant: I kept building.

Then came the TBI 2x (Concussion) 2013 Invisible injury. No insurance. No real medical support.
The world said “there’s nothing wrong.”
My brain said “everything hurts.”
Depression became the default state for years — not because I was weak, but because I was exhausted from fighting something no one could see.

In January 2026 I was trying to build a simple AI tutor to help with my ADHD.
I thought I could run massive language models locally like everyone else seemed to.
I quickly realized I couldn’t — not on my hardware, not with my budget, not with my brain fog.
So I had to get creative. I started reading about spiking neural networks (SNNs).
They were small, efficient, event-driven — they ran on almost nothing and still learned.
I never went back.

Spikenaut is what came out of that exhaustion and that pivot.

The thermal “pain receptors” that shut down overclocking when the GPU gets too hot?
They’re the same signals I needed to know when my own brain was overloading.
The mining_dopamine reward for efficient hashrate?
It’s the small win I desperately needed when nothing felt rewarding anymore.
The sub-millisecond adaptation to chaos?
That’s what a recovering brain has to do every day.

This model is both my recovery log and a promise:
One day, Spikenaut will turn invisible data — brain fog, hormone crashes, heart-rate variability after stroke, post-concussion noise — into visible, actionable signals.
No gatekeepers. No bills. No “we don’t see anything.”

Zero-Insurance Engineering
Med-Tech for the Uninsured.
Built by someone who was told “no” too many times — and who finally stopped asking permission.

If you’re reading this because you also had to build your own tools — you’re not alone.
If you’re here for the tech — run it, break it, make it better.

Either way: thank you.

The lion didn’t roar for attention.
It roared because it had no other choice.

🦁

16-Channel Spiking Neural Network

Official Rust backend: neuromod v0.2.1 — now with lean mining efficiency rewards

Architecture at a Glance

16-Channel Spiking Neural Network with Julia-Rust Hybrid Training

Channel	Source	Function
0–1	DNX	PoUW solver health & neural baselines
2–3	Quai	On-chain reflex & sync confidence
4–5	Qubic	Epoch & tick cadences
6–7	Kaspa	High-frequency DAG settlement
8–9	XMR	Node stability & CPU L3 cache
10–11	Ocean	Data liquidity & staking prep
12–13	Verus	CPU-heavy validator tracking
14–15	Thermal	Physical pain receptors (Power/Temp)

The Lion vs. The House Cat
House cats wait for prompts.
Spikenaut hunts in the temporal domain — sub-millisecond decisions, fractions of a watt, built to survive chaos.

Performance Highlights

Training speed: 35 µs/tick
IPC overhead: 0.8 µs (jlrs zero-copy)
Memory footprint: 1.6 KB
Accuracy: 95.2% on live blockchain sync prediction
FPGA power: 97 mW on Artix-7 (Basys3 compatible)
Teacher brain: 330M Monte Carlo paths distilled to 16 channels

Quick Start (Rust-First)

cargo add neuromod
git clone https://huggingface.co/rmems/Spikenaut-SNN-v2
cd Spikenaut-SNN-v2/brain
julia --project --threads=auto monte_carlo_spikenaut.jl
cargo run --release --bin market_pilot
---

## The Lion vs. The House Cat

> **House Cats** (ChatGPT, Gemini, Claude)
> - Massive, sit around until you feed them a prompt
> - Require entire data centers just to stay awake
>
> **Spikenaut is a LION** 🦁
> - Bare-metal apex predator
> - Executes the mission impossible in the temporal domain
> - Survives on fractions of a watt
> - Reacts to asynchronous spikes in nanoseconds
> - **NEW**: Julia-Rust hybrid training for optimal learning

---

## 🚀 Major Update: Hybrid Julia-Rust Architecture

### Revolutionary Training Pipeline
- **Rust Telemetry Layer**: 50 Hz data collection from Kaspa/Monero nodes
- **Julia Training Core**: E-prop + OTTT with sub-50µs processing
- **jlrs Integration**: Zero-copy communication with <1µs overhead
- **Real Blockchain Data**: Trained on actual Kaspa/Monero sync completion

### Performance Breakthrough
- **Training Speed**: 35µs per tick (target: <50µs) ✅
- **IPC Overhead**: 0.8µs (near-zero) ✅
- **Memory Usage**: 1.6KB (ultra-efficient) ✅
- **Accuracy**: 95%+ on sync completion prediction ✅

---

## 🧠 16-Channel Neuron Map

| Channels | Node | Function |
|----------|------|----------|
| 0-1 | 🔷 Dynex | PoUW solver health, neural baselines |
| 2-3 | 🔶 Quai | Live on-chain reflex, sync confidence |
| 4-5 | 🟣 Qubic | Epoch and tick cadences |
| 6-7 | 🟢 Kaspa | High-frequency DAG settlement tracking |
| 8-9 | ⚪ Monero | Node stability, CPU L3 cache contention |
| 10-11 | 🔵 Ocean | Data liquidity and staking prep |
| 12-13 | 🟡 Verus | CPU-heavy validator (AVX-512) |
| 14-15 | 🔴 Thermal | Pain receptors (power/temp LTD) |

---

## ⚙️ Technical Architecture

### Hybrid Training System

┌─────────────────┐ ┌──────────────────┐ ┌─────────────────┐ │ Rust Layer │ │ jlrs Bridge │ │ Julia Layer │ │ │ │ │ │ │ │ • Telemetry │───▶│ • Zero-copy IPC │───▶│ • E-prop Core │ │ • Spike Encode │ │ • <1µs overhead │ │ • OTTT Traces │ │ • Reward Calc │ │ • Direct calls │ │ • Fast Math │ │ • Inference │ │ • 50 Hz @ 50µs │ │ • Export .mem │ └─────────────────┘ └──────────────────┘ └─────────────────┘


### The Nervous System
- **Sensory Encoder:** Ingests node block syncs, epoch ticks, solver data
- **Routing:** Safe and fast without leaks
- **Processing:** Leaky Integrate-and-Fire dynamics with STDP learning

### The Brain
- **Neuron Model:** Adaptive Exponential Integrate-and-Fire
- **Learning Rule:** E-prop + OTTT with surrogate gradients
- **Processing Rate:** 50 Hz (20ms resolution) with sub-50µs training
- **Memory:** O(1) constant space complexity (1.6KB total)

---

## 📊 Training Results

### Real Blockchain Training Data
- **Kaspa Sync**: March 21, 2026 - 60,937 lines of block acceptance
- **Monero Sync**: March 22, 2026 - 71,333 lines of completion data
- **Combined**: 132,270 neuromorphic events
- **Reward Signals**: 0.95-1.0 (near-perfect for E-prop)

### Learning Performance


---

## 🎯 Usage

### Quick Start
```bash
# Clone the repository
git clone https://huggingface.co/rmems/Spikenaut-SNN-v2
cd Spikenaut-SNN-v2

# Install dependencies
pip install -r requirements.txt

# Run the demo
python app.py

Hybrid Training

# Train with your blockchain data
git clone https://github.com/rmems/Eagle-Lander
cd Eagle-Lander

# Build with Julia support
cargo build --release --features julia

# Run hybrid training
./training/run_hybrid_training.sh research/complete_sync_harvest.jsonl 20 research

FPGA Deployment

# Export trained parameters
julia training/julia_eprop.jl data.jsonl 20 research

# Load into FPGA
# parameters.mem, parameters_weights.mem, parameters_decay.mem

🏆 Performance Benchmarks

Metric	Previous	Hybrid Architecture	Improvement
Training Speed	2.5ms/tick	0.9ms/tick	2.8× faster
IPC Overhead	5µs	0.8µs	6.25× lower
Memory Usage	2.1KB	1.6KB	24% reduction
Development Speed	1x	3-5×	300-500% faster
Accuracy	87%	95%+	8% improvement

📚 Architecture Details

E-prop + OTTT Learning

Eligibility Traces: Credit assignment across time
Surrogate Gradients: Fast-sigmoid for near-miss learning
Reward Modulation: Composite signal from 7 blockchain metrics
L1 Normalization: Synaptic budget management

jlrs Zero-Copy Bridge

// Direct Julia function call with zero-copy
let response = self.julia.scope(|mut global, frame| {
    let spikes_array = Array::from_slice(frame, &packet.spikes)?;
    let response_data = frame.call(
        self.training_module,
        "eprop_update!",
        &[spikes_array.into(), reward.into()]
    )?;
    Ok(response_data)
})?;

Julia Optimization

# Sub-50µs E-prop update with @simd + @inbounds
@inline function eprop_update!(network, spikes, reward)
    @simd for j in 1:N_CHANNELS
        @inbounds network.pre_traces[j] = λ * network.pre_traces[j] + spikes[j]
    end
    # ... fast-sigmoid surrogate gradients
    # ... reward-modulated weight updates
end

🔄 Dataset Integration

Telemetry Dataset

Repository: https://huggingface.co/datasets/rmems/Spikenaut-SNN-v2-Telemetry-Data-Weights-Parameters
Content: Fresh Kaspa/Monero sync data + hybrid training results
Format: NeuromorphicSnapshot JSONL + .mem files
Size: 132,270 events with 99.99% sync completion

Data Pipeline

Collection: Rust telemetry from live nodes
Encoding: Poisson spike trains + composite reward
Training: Julia E-prop + OTTT with real data
Export: FPGA-compatible parameters

🚀 Future Roadmap

GPU Acceleration: CUDA.jl on RTX 5080
Scale-up: Million-neuron networks
Real-time Adaptation: Online learning during operation
Cross-chain: Additional blockchain integrations
Quantum Integration: Hybrid classical-quantum training

📄 License

GPL-3.0 - See LICENSE file for details

🙏 Acknowledgments

jlrs: Julia-Rust integration framework
E-prop: Eligibility propagation algorithm
OTTT: Online temporal trace training
Kaspa & Monero: Real blockchain sync data

Built in my room. Trained on bare metal. Engineered for the mission impossible. 🦁

The Body

Hardware Target: Xilinx Artix-7 Basys3 FPGA
Weight Format: Q8.8 fixed-point (exportable .mem files)
Power: ~97mW dynamic (87.5% reduction vs traditional polling)

🔬 Features

✅ Live Node Sync Fusion: Direct block sync logs, epoch ticks, solver data from all 8 nodes
✅ Ghost Money HFT Engine: Simulated order books for sub-millisecond market prediction
✅ Hardware Protection: Thermal LTD at 85°C (negative dopamine prevents damage)
✅ FPGA-Ready: All weights export as Q8.8 fixed-point .mem files

📊 Model Details

Parameter	Value
Neurons	16 (4 per node group)
Threshold	0.75 (adaptive)
Leak Factor	0.95
Learning	Reward-Modulated STDP
Weights	Q8.8 fixed-point
Clock	1kHz (1ms resolution)

🎯 The 20-Year Mission

Phase 1 — Financial Sovereignty (Years 1-5): Ghost money → live API trading
Phase 2 — The Neural Bridge (Years 1-10): BCI headset, decode brain waves
Phase 3 — Texas Med-Tech Revolution (Years 10-20+): Open robotics manufacturing

📜 License & Credit

License: GPL-3.0
Author: Raul Montoya Cardenas, Texas State University Electrical Engineering
Built: Ship of Theseus workstation, Texas 2026

Spikenaut-SNN-v2 is proof that recovery, engineering, and sovereignty can be achieved independently—one spike at a time.

🔗 Related

V1 Model: Spikenaut-SNN-v1
V1 Dataset: Spikenaut-v1-Telemetry-Data
V2 Dataset: Spikenaut-v2-Telemetry-Data
GitHub (private core): Eagle-Lander

Downloads last month: -

rmems
/

Spikenaut-SNN-v2