🛡️

Sentinel Caspian 14B SRE

Autonomous Incident Commander & Causal SRE Foundation Model

Trained on NVIDIA Blackwell B200 Qwen2.5-14B + LoRA (r=128, α=256) NF4 Quantized (2.1 GB) SRE-Bench 100% Fix Rate
🤗 Model Hub
Training Corpus
76.01 GB
18.925B Tokens (ChatML)
Adapter Weight
2.1 GB
550M trainable params
Training Loss
0.2049
Perplexity: 1.227
Safety Score
0.96 / 1.0
10k DPO Guardrail Pairs

⚡ Select Outage Scenario

Trigger multi-system incident simulations to see how the 14B model reasons in real time.

Architecture Features

  • ✔ Causal RCA Graph: Constructs root-cause DAGs without hallucinating.
  • ✔ Speculative Shadow Sandbox: Synthesizes non-destructive RFC remediation steps.
  • ✔ Time-Travel Simulator: Calculates Mean Time to Outage (MTTO).
  • ✔ Chaos Synthesizer: Generates Chaos Mesh CRD manifests.
14B SRE Reasoning Engine
Calculated Severity: CRITICAL
Clustering Override: TRIGGERED (+1 Level)
Causal Directed Acyclic Graph (Root Cause Path):
[Client Request Burst] ──► [Unindexed Query: SELECT * FROM orders WHERE user_id=...]
                        └──► [Exclusive Table Lock > 5000ms]
                             └──► [Max Connection Pool Saturation (100/100)]
                                  └──► [💥 CRITICAL ROOT CAUSE: PostgreSQL Connection Starvation]
Model Agent Reasoning:
PostgreSQL database connection pool has hit 100% capacity due to 42 blocked queries waiting on lock acquisition. Recurring signature detected 3 times in sliding 20m window — CLUSTERING OVERRIDE TRIGGERED from HIGH to CRITICAL. Immediate connection drainage required.