Explore the projects, workshops, and tutorials that Graziano Casto has created.

Projects - Graziano Casto

A closed-loop Kubernetes resource right-sizing system that observes workload metrics, evaluates them via OPA policies, and opens a GitOps PR with updated CPU/memory manifests — no direct cluster patching.

A closed-loop Kubernetes resource right-sizing system built around a two-layer OPA policy architecture: a master engine managed by the platform team and per-workload Rego files owned by SREs. The orchestrator queries Prometheus for P95 CPU and peak memory data (with native JVM heap/non-heap metrics via OTel Java agent), sends only observational data to OPA, and — when a resize is recommended — creates a branch, patches the manifest using ruamel.yaml to preserve comments and formatting, and opens a GitHub PR with a full diagnostics table. Human-in-the-loop by design; full automation is possible via Renovate or Mergify.

K8s Resource Optimizer

A production-realistic observability pipeline for constrained edge environments, featuring tail-based sampling, file-backed persistent queues, unified signal correlation, and automatic gap-fill after link restoration.

A production-realistic observability pipeline designed for constrained edge environments with intermittent connectivity and limited resources. The edge node runs a demo Go app alongside a custom OTel Collector with tail-based sampling (buffers full traces before deciding keep/drop, retaining 100% of errors and high-latency traces while dropping ~80% of fast successful spans) and a file-backed bbolt queue providing ~47 hours of outage tolerance. Fluent Bit with Lua filtering applies identical sampling criteria to logs, achieving ~86% volume reduction while maintaining 1:1 alignment with kept traces. Every log carries a trace_id for cross-signal correlation in Grafana. On link restore, both Loki and Jaeger recover queued telemetry with original timestamps. Deployable locally via k3d or on Civo managed Kubernetes with no application changes.

Observability on Edge

Extract, validate, analyze & version control your entire Kubernetes cluster. Zero-config backup, comprehensive reporting & Git integration.

KALCO is a comprehensive Kubernetes cluster management tool designed for Platform Engineers who need to extract, validate, and version control their entire cluster configuration. It provides zero-config backup capabilities, comprehensive reporting, and seamless Git integration. With KALCO, you can track changes to your cluster configuration over time, ensuring compliance and enabling easy rollbacks when needed.

KALCO

 A Google ADK voice agent implementation using Gemini 2.5 Flash with Google Search, Filesystem access and PDF reading capabilities.

Jarvis ADK Voice Agent is an advanced implementation of Google's ADK that leverages Gemini 2.5 Flash for intelligent voice interactions. This project demonstrates modern AI capabilities including Google Search integration, file system access, and PDF reading functionality. Perfect for developers looking to build sophisticated voice-controlled applications with cutting-edge AI technology.

Jarvis ADK Voice Agent

Tutorial for deploying LLMs on local Kubernetes (Minikube on Apple Silicon w/ GPU) following MLOps (GitOps) best practices with Helm, ArgoCD, and Istio.

A comprehensive tutorial repository for deploying Large Language Models (LLMs) on local Kubernetes environments, specifically optimized for Minikube on Apple Silicon with GPU support. This project follows MLOps and GitOps best practices, demonstrating the integration of Helm for package management, ArgoCD for continuous deployment, and Istio for service mesh capabilities. Perfect for developers learning to productionize GenAI models.

Deploying GenAI Models on Kubernetes

A collection of tutorials to create ephemeral environments with crossplane e kube-green.

A collection of practical tutorials demonstrating how to create and manage ephemeral Kubernetes environments using Crossplane for Infrastructure as Code and kube-green for automatic resource scheduling. This project showcases efficient resource utilization patterns, demonstrating how to automatically spin up and tear down development environments based on schedules, reducing infrastructure costs while maintaining development productivity.