Building SOC-AI: Sub-Second Security Triage with Groq & Kestra

Quantum Sage — Tue, 06 Jan 2026 12:38:00 +0000

The Motivation: Solving Alert Fatigue 🛡️

Security teams are drowning in logs. I built SOC-AI solo during the AI Agents Assemble Hackathon (hosted on @wemakedevs) to prove that AI can handle the "boring" triage while humans keep control.

The Architecture

Triage: Groq (Llama-3.3) for 500ms log analysis.
Orchestration: Kestra for semi-autonomous remediation.
Frontend: Next.js 15 (Live on Vercel!).

Technical Deep Dive – Reliable AI & Orchestration

One of the biggest challenges in AI agents is hallucination. In a security context, a "hallucinated" IP address or action could be a disaster. I solved this using two core patterns:

1. Forcing Structured Output with Zod & Groq

I didn't just ask the AI to "analyze the log." I defined a strict contract using Zod. This ensures that the high-speed Llama-3.3 model on Groq returns a precise JSON object that my backend can trust.

// Define the Triage Schema
const SecurityTriageSchema = z.object({
  severity: z.enum(["low", "medium", "high", "critical"]),
  threat_type: z.string().describe("Categorization like Brute Force, SQLi, etc."),
  action_suggested: z.enum(["block_ip", "disable_user", "monitor"]),
  reasoning: z.string()
});

// Forcing Groq to adhere to the schema
const chatCompletion = await groq.chat.completions.create({
  model: "llama-3.3-70b-versatile",
  response_format: { type: "json_object" }, // Crucial for reliable JSON
  messages: [
    { role: "system", content: "You are a SOC Triage Agent. Output ONLY JSON." },
    { role: "user", content: `Analyze this log: ${rawLogData}` }
  ],
});

2. Semi-Autonomous Remediation via Kestra

Trust is everything in security. Instead of letting the AI run wild, I built a Human-in-the-Loop system.

When the AI suggests an action (like block_ip), it appears on my Next.js Dashboard. Only after I click "Approve" does the backend trigger a Kestra Workflow.

Why Kestra?

Retries: If the Firewall API is down, Kestra handles the retry logic.
Audit Trail: Every action taken is logged visually in the Kestra UI.
Separation of Concerns: My Next.js app handles the UI, while Kestra handles the heavy infrastructure automation.

The Result

By combining Groq's speed with Kestra's reliability, SOC-AI can triage a log and present a remediation plan to an analyst in less than a second.

This project was a solo build, and it taught me that the future of AI isn't just about the "chat" - it's about orchestration.

Links & Demo

GitHub: https://github.com/kaushik0010/soc-ai
Video Walkthrough: https://youtu.be/LbuHXiPznJE

What are your thoughts on semi-autonomous security? Would you trust an AI to suggest your firewall rules? Let's talk in the comments!

DEV Community: Quantum Sage