Upload folder using huggingface_hub

README.md

@@ -2,6 +2,9 @@
 license: mit
 tags:
   - networking
+  - sase
+  - agentic
+  - autonomous-agents
   - congestion-control
   - loss-prediction
   - rust
@@ -16,12 +19,33 @@ pipeline_tag: other
 
 # 🦀 Oxidize ML Models
 
-Machine learning models for [Oxidize](https://github.com/gagansuie/oxidize) — an Agentic SASE Platform built in pure Rust.
+Machine learning models for [Oxidize](https://github.com/gagansuie/oxidize) — an **Agentic SASE Platform** built in pure Rust.
 
-> Neural networks predict packet loss before it happens, optimize routing in real-time, and accelerate your network automatically.
+> Oxidize doesn't just accelerate your network — it thinks for you. Autonomous agents predict, decide, and act in real-time to optimize every connection.
+
+## From Network Accelerator to Agentic SASE
+
+Oxidize has evolved beyond traditional network acceleration. Here's what makes it different:
+
+| Traditional Network Accelerator | Oxidize Agentic SASE |
+|--------------------------------|---------------------|
+| Static routing | **Autonomous path selection** per flow |
+| Reactive congestion control | **Predictive agents** that act before congestion |
+| Manual policy updates | **Self-healing** that detects and responds to threats |
+| One-size-fits-all | **Per-flow bandits** that learn your traffic patterns |
+
+### The Agentic Difference
+
+Oxidize runs **autonomous prediction-action agents** in the packet path:
+
+1. **Per-Flow UCB1 Bandits** — Each network flow gets its own bandit that learns optimal routing decisions over time
+2. **Online Model Distillation** — New models validated in shadow mode, atomic swap into production without downtime
+3. **Self-Healing** — Anomaly detection tracks validation failures; triggers automatic model retraining when >50% of samples are rejected (e.g., during DDoS attacks)
+4. **Continuous Learning** — Real traffic from your servers improves the models daily
 
 ## Highlights
 
+- **Autonomous Agents** — No manual intervention required; agents make decisions in microseconds
 - **10x faster inference** via INT8 quantization
 - **<1µs cached latency** with speculative pre-computation
 - **Pure Rust** — no Python runtime, trained with Candle

config.json

@@ -1,7 +1,7 @@
 {
   "metadata": {
     "contributing_servers": 1,
-    "last_updated": "2026-03-03",
+    "last_updated": "2026-03-04",
     "total_samples": 20000
   },
   "models": {

ppo_congestion.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fadef7def707d2a646b79ff3c1e99a1c2593e2f58c7303958791362272fb64e1
+oid sha256:cbbfedf70cbdf5135a06b017501adcd3f04685f66269b0786d4c10f43d3e7fce
 size 143796

transformer_loss.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:778c7306e99e3828f3f28fddf2ed8e3552a6418a99df7e6133fd984c8058bb76
+oid sha256:4668504c7b2608120f1b06c1a631419154a3e0f4f7ff5a038b667e510b8750bd
 size 199908