I automated myself out of my own workflow
I kept doing the same infrastructure tasks every week. Same checks, same provisioning, same cleanup. So I built agents to do them instead.
Hi, I'm @navrast.
Architect and leader in daylight. Dirty-handed owl engineer at night. I design AI infrastructure at national scale and come home and build the tools I wish existed. Everything I make, I use.
Right now I'm exploring multi-agent robotics, building an air-gapped home AI, running local models on a Mac Studio cluster, and shipping OpenDray, a terminal cockpit for AI coding agents.
Posts
The weekend hack that became my daily driver
I added a Telegram bridge to OpenDray so I could check on Claude sessions from my phone. It was supposed to take a Saturday.
Everyone talks about training. Nobody talks about the cooling.
I was designing liquid-cooled GPU processing farms 14 years ago. The conversation has changed. The physics hasn't.
I wanted the best price. So I built a crawler.
ShopSerp watches 250+ stores across 17 countries. It started because I was shopping for a camera lens.
My lab is not a hobby
Containers, orchestration clusters, multiple cloud regions, a mesh network tying it all together. I refuse to recommend architecture I haven't broken myself.
What happens when you give AI agents your cloud credentials
You need a kill switch that doesn't depend on you being awake.
No Docker. No Nginx. No drama.
go:embed bakes the Flutter web build into the Go binary. One file to copy, one process to run.
Selected work
FleetBay 2025
Multi-agentic warehouse and logistics management. Shipping, inventory, returns, barcode scanning, serial tracking, vendor workflows. Built for a friend whose business ran on spreadsheets.
DynaCore 2025
Multi-agent orchestration framework. 14 AI agents, each one a domain specialist. They handle cloud provisioning, identity management, network configuration, monitoring, and code delivery. I handle decisions.
Operantis 2026
Framework for building AI employees for any business. Define the domain rules, connect the tools, set the guardrails. The agents handle the repetitive work. Humans handle the exceptions.
ShopSerp 2026
Agentic market intelligence. Crawls 250+ stores across 17 countries, tracks prices, finds the best deals. Started because I was tired of checking five websites for camera gear.
OpenDray 2026
Self-hosted terminal cockpit for AI coding agents. Run Claude, Codex, Gemini from your phone. Telegram bridge for remote control. Open plugin standard. Go + Flutter + PostgreSQL.
About
I've spent most of my career designing infrastructure for large organisations. Storage platforms, hybrid cloud, AI compute. The kind of systems where a wrong decision costs millions and a right one runs for a decade.
Somewhere along the way I started building things on the side. Not because I had to. Because I'd see a problem at work, think about it on the drive home, and by midnight I'd have a prototype running. A warehouse platform because a friend's business was drowning in spreadsheets. A price tracker because I got tired of overpaying for camera gear. An agent framework because I realised I was doing the same infrastructure tasks every week and kept wondering why I hadn't automated myself out of the loop.
That last one changed how I work. I now run a multi-agent orchestration framework where AI agents handle the operational work I used to do manually. Cloud provisioning, identity management, network configuration, monitoring, code delivery. The goal is full hands-off on the routine stuff so I can spend my time discovering what's next instead of maintaining what already exists.
I was designing modular GPU-clustered liquid-cooled processing farms 14 years ago, before anyone was calling it "AI infrastructure." Today I design AI compute platforms at national scale. Fabric topology, storage tiering, inference pipelines, cooling envelopes. The hardware changed. The thinking didn't.
I don't just call APIs. I run models locally on a cluster of Mac Studios, test their limits, build evaluation datasets, and try to understand why they fail before I trust them with anything real. When a model gets something wrong I want to know if it's the data, the quantisation, or the prompt.
I run a scaled hybrid lab. Containers, orchestration clusters, multiple cloud regions, all meshed and secured end to end. It's where I test architectures before I recommend them to anyone. If I'm going to stand in a steering committee and say something works, I want to have broken it first.
At home I'm building something I've wanted since I was a kid. An AI that lives inside the house. It sees, it listens, it understands context, it controls the environment. Not a smart speaker with a wake word. A presence that follows what's happening and responds naturally. And it stays inside. No cloud. No external exposure. Private by architecture, not by policy.
I don't really have a specialty. I have a pattern. See a problem. Understand it deeply. Build a system that solves it. Then build an agent that runs the system. The domain changes. The approach doesn't.
I get genuinely excited when things click. When a small model running locally extracts business rules from a photo of a receipt. When a single Go binary embeds an entire mobile web app. When a Telegram bot renders an AI agent's permission dialog as inline keyboards and you approve a code commit from the bus.
Those moments are why I still build things after 25 years.