Speculative-Guard AI

An autonomous self-healing DevOps agent that detects cloud server crashes, searches GitHub/docs via Tavily, and deploys verified code patches using Nebius AI Cloud.

Comment

# 🛡️ Speculative-Guard AI

Autonomous Self-Healing Cloud & Infrastructure Agent powered by Nebius AI Cloud & NVIDIA Open-Source Models.

Inspiration

In modern cloud computing, infrastructure downtime and unhandled runtime errors cost global enterprises billions of dollars annually. When a critical production server crashes at 3:00 AM, engineers spend hours parsing complex stack traces, searching through StackOverflow, and manually debugging code. We were inspired by a simple question: What if servers could heal themselves in real time before engineers even wake up? By combining ultra-fast open infrastructure with agentic workflows, we set out to build a fully automated DevOps responder.

What it does

Speculative-Guard AI is an autonomous self-healing agent that continuously monitors software server logs. When a crash or stack trace occurs, the system automatically:

  1. Parses the Log: Uses NVIDIA open-weights models hosted on Nebius Token Factory to isolate the root cause.
  2. Researches the Web: Dynamically queries GitHub issues, StackOverflow, and library documentation via Tavily API to find official patches.
  3. Generates & Tests Code: Automatically writes a corrected code patch and runs it inside an isolated sandbox verification environment to verify zero regressions before deployment.
  4. Streams Telemetry: Displays the entire reasoning and repair loop live on a dark-mode Next.js DevOps dashboard.

How we built it

We built the application using a robust, 100% open-source tech stack:

  • Backend & Agent Orchestration: Python, FastAPI, and modular LangGraph control loops.
  • AI Inference: NVIDIA Meta-Llama 3.1 open-weights models served through high-throughput Nebius AI Token Factory API endpoints.
  • Live Web Research: Tavily API for real-time developer document and patch retrieval.
  • Sandbox Verification: Local Docker container environments for safe test-patch execution.
  • Frontend UI: Next.js, TypeScript, and Tailwind CSS for the real-time DevOps command center.

Challenges we ran into

  • Hallucination Prevention: Ensuring the LLM didn't invent non-existent APIs required strict prompt constraints and rich context injection via Tavily search results.
  • Sandbox Security: Safely running generated code patches required implementing execution timeouts and isolated temporary runtimes to prevent host system leakage.
  • Latency Optimization: Chaining multiple agent calls can slow down response times. Using Nebius Token Factory gave us the high throughput required for near-instant execution.

Accomplishments that we're proud of

  • Sub-15 Second Resolution Times: Successfully reduced Mean Time to Resolution ($\text{MTTR}$) from roughly 45 minutes down to under 15 seconds.
  • True Agentic Workflow: Built a complete, autonomous loop (Perception $\rightarrow$ Research $\rightarrow$ Execution $\rightarrow$ Verification) without human intervention.
  • Mathematical Efficiency: Proved an overall workflow efficiency gain of approximately $99.44\%$ over human manual debugging cycles.

What we learned

  • Open-Source Infrastructure Mastery: Hosting open weights like Llama 3 on Nebius Cloud provides flexibility and performance matching proprietary closed APIs.
  • Agentic Research Integration: Integrating Tavily directly into LLM decision loops dramatically improves bug resolution accuracy compared to standalone models.

What's next for Speculative-Guard AI

  • Multi-Container Fleet Management: Expanding support from single-service scripts to multi-container Kubernetes clusters.
  • Enterprise Git Integration: Automatically opening, testing, and merging pull requests directly into corporate GitHub and GitLab repositories.
  • Advanced Predictive Healing: Utilizing historical telemetry data to predict server crashes before they occur.

Built With

Share this project:

Updates

Submission history