Sentinel — AI Incident Response Copilot

Every engineering team faces the same problem: logs and alerts flood in constantly, and someone has to manually judge what's actually serious. That triage step wastes time, distracts engineers from real work, and doesn't scale as systems grow more complex. Sentinel automates this first, most tedious step of incident response using a hybrid AI architecture designed to balance cost and accuracy. At its core is a Gemma 2 model that I personally fine-tuned using LoRA on an AMD Instinct MI300X GPU. This local model handles the vast majority of incoming logs instantly and completely free of charge, classifying severity, category, and providing a short explanation. Only when this local model's confidence is low does Sentinel escalate the log to a larger, more capable reasoning model — Kimi K2.7 Code — through the Fireworks AI API. Every single classification, along with which model handled it, is saved to a database and displayed on a live dashboard. This gives teams a direct, measurable view of their cost savings: exactly what percentage of incidents were resolved for free locally versus how many required the more expensive remote model. Unlike many AI hackathon projects that simply wrap a single API call in a chat interface, Sentinel's core innovation is its routing decision layer — a genuinely fine-tuned model making real-time, defensible tradeoffs between cost and accuracy. This entire system, from model fine-tuning to inference to the user-facing dashboard, runs on AMD compute, and is fully containerized with Docker for easy deployment. This isn't just a hackathon prototype — the same architecture could plug directly into real-world alerting pipelines like PagerDuty, Datadog, or Slack integrations, immediately reducing both AI inference costs and mean-time-to-triage for engineering teams of any size. It solves a problem every company running infrastructure genuinely has.