GenKube Guard

🚀 Inspiration

I never imagined I’d be submitting to a global hackathon solo, with just my basic laptop and zero prior hackathon experience.

It all started during an internship at a reputed tech firm in India. I was given real tasks but no mentorship. No one to ask for help. At first, I was terrified. I didn't even know where to begin. But I forced myself to sit down, day after day, and figure things out alone.

That struggle became my strength.

Later, while working on a side project, I found myself constantly battling YAML files. So I built a basic YAML generator just for myself. That project opened my eyes. What if I could go beyond just generating YAMLs? What if I could analyze and patch them intelligently using AI?

That idea planted the seed for GenKube Guard and even though I was scared of failing, I knew I had to try.

🔧 What it does

GenKube Guard is an LLM-powered DevSecOps assistant for Kubernetes YAMLs. It:

Analyzes YAMLs using kube-linter
Explains issues in plain language using an LLM
Auto-patches insecure Deployments and StatefulSets
Provides persona-based suggestions (junior, senior, SRE)
Integrates cultural DevSecOps recommendations using a mock Qloo API
Enables memory search via GraphQL and FAISS RAG

🏗️ How I built it

Backend: FastAPI with routes like /analyze, /patch, /suggest, /recommend, /memory
LLM: Mistral via Ollama (run locally)
Mock Qloo Integration: Built fallback response logic when the API key didn’t work
Memory Engine: FAISS-based vector store with semantic fallback
Prompt Engineering: Custom YAML-aware templates for each endpoint
Deployment: Shifted to Hugging Face when Render free tier maxed out (not compatible with Ollama)

🧱 Challenges I ran into

LLM inference was very slow especially on a basic 16GB RAM laptop
Docker rebuilds took forever, and sometimes froze the system
Qloo API didn’t respond, so I manually crafted fallback responses for every query
Couldn’t deploy LLM backend live on Hugging Face (Ollama is local only)
Most difficult of all: emotional exhaustion. I cried some nights, doubting myself, wondering if I could finish.

But something kept me going maybe it was the thought that I’ve come this far, there’s no turning back.

🏆 Accomplishments that I'm proud of

Built GenKube Guard completely solo design, code, testing, and deployment
Started from a simple YAML generator → ended with an LLM-powered assistant
Every endpoint was tested with multiple YAMLs and edge cases
Created persona-aware, culturally tuned responses without real Qloo access
Pushed past fear, burnout, and system limitations and finished strong

📚 What I learned

Kubernetes internals, YAML structuring, kube-linter
FastAPI, Docker, Ollama, FAISS, RAG, and GraphQL
How to write structured prompt templates for secure AI output
Resilience: the ability to keep going even when nothing seems to work
That done with heart beats perfect with fear

🔮 What’s next for GenKube Guard

Cloud-hosted LLM inference to enable full production deployment
Full Qloo API integration (once accessible)
A lightweight web UI (maybe React or plain HTML)
DevSecOps CI integration (e.g., GitHub Actions plugin)
Open-sourcing the project and documenting it to help others build YAML tooling with LLMs

💔 The Last-Minute Challenge

Right before submitting, Render’s free tier maxed out.
I rushed to deploy on Hugging Face Spaces, but there I hit the final roadblock:

Ollama doesn’t run on Hugging Face — meaning my deployed app can’t generate real LLM responses.

So for /recommend and /suggest, the app only returns mock fallback data.
Yes, it breaks part of the magic.
Yes, I was devastated.
But I’ve shared my full local demo video — where everything works perfectly: real prompts, real YAMLs, real LLM inference.

🧡 Thank you for reading.

This isn’t just a tool. It’s the story of how I stopped being afraid and started building.

– Aswathi VK