gagansuie/oxidize-models

🦀 Oxidize ML Models

Machine learning models for Oxidize - an enterprise-grade network backbone built in Rust.

Models

Tier 1 - Core Intelligence

Model	Architecture	Purpose	Input	Output
transformer_loss	Transformer (d=64, 4 heads)	Predict packet loss 50-100ms ahead	20×8 sequence	Loss probability
ppo_congestion	PPO Actor-Critic (128 hidden)	Optimize congestion window	8-dim state	Continuous CWND

Tier 2 - Advanced Optimization

Model	Architecture	Purpose	Input	Output
compression_oracle	MLP classifier	Decide optimal compression strategy	8 features	4 classes
path_selector	Contextual bandit	Select best path per traffic type	29 features	4 paths

Architecture

Transformer Loss Predictor

Input: [batch, 20, 8]  →  MultiHeadAttention(d=64, h=4)  →  FFN  →  Linear(1)  →  Sigmoid  →  Loss probability

Features (8):

• RTT (normalized)
• RTT variance (jitter)
• Bandwidth estimate
• Current loss rate
• Packets in flight
• Buffer occupancy
• Inter-packet gap
• Time since last loss

PPO Congestion Controller

Input: [batch, 8]  →  Actor(128)  →  ReLU  →  Actor(128)  →  ReLU  →  (mean, log_std)  →  CWND multiplier

Actions (6):

• DecreaseLarge (-25% CWND)
• DecreaseSmall (-10% CWND)
• Maintain (0%)
• IncreaseSmall (+5% CWND)
• IncreaseAdditive (+1 MSS)
• IncreaseLarge (+10% CWND)

Usage

With Oxidize (Rust)

use oxidize_common::model_hub::{ModelHub, HubConfig};

let hub = ModelHub::default_config();
let paths = hub.download_models()?;

// Models auto-loaded into ML engine

Environment Variables

# Optional: For private repos or uploads
export HF_TOKEN=hf_xxxxxxxxxx

Training

Models are trained on real network telemetry from Oxidize deployments:

1. Data Collection: Servers collect LossSample and DrlExperience during operation
2. Aggregation: Training data uploaded to this repo's training_data/ folder
3. Training: Candle-based training in pure Rust
4. Deployment: Updated models pushed here, servers auto-sync

Contributing Training Data

Oxidize servers can opt-in to contribute anonymized training data:

let config = HubConfig {
    upload_training_data: true,
    ..Default::default()
};

License

MIT - Same as Oxidize

Citation

@software{oxidize2026,
  author = {gagansuie},
  title = {Oxidize: Enterprise-grade network backbone},
  url = {https://github.com/gagansuie/oxidize},
  year = {2026}
}