Loading…
Loading…
Journal
Short write-ups on the decisions and trade-offs behind the systems I build.
What everyone "knows": AI makes us efficient. We build systems faster, cleaner, smarter — so surely we'll use fewer resources, not more. What actually happens: cheaper never means less. Make creation…
What everyone "knows": locking down the network is free security. Deny all outbound traffic by default, allow only what you need. Pure upside — what's the downside of being careful? What actually…
What everyone "knows": securing an AI endpoint is one thing — check an API key. Key's good, you're in. What actually happens: "who are you?" has several universes' worth of answers. Humans and robots…
What everyone "knows": connecting my gateway to an external service is just routing. I have the URL and a token — what could possibly go wrong? What actually happens: "the network reaches it" is the…
What everyone "knows": a gateway is a glorified signpost. It forwards your request to a backend and gets out of the way. Maybe it adds a little latency. What actually happens: on a single hop, the AI…
What everyone "knows": I run three control-plane nodes, so my cluster is highly available. I can replace them like any other server — they're cattle, not pets. What actually happens: three is the…
What everyone "knows": to set up an AI coding assistant for a whole team, you write a setup guide, paste a config file in Slack, and spend the next two weeks answering "it doesn't work on my…
What everyone "knows": put an AI reviewer on your pull requests and it becomes the gatekeeper — a fancy linter that blocks the merge until it's happy. What actually happens: the reviewer can do far…
What everyone "knows": the AI writes the code, you run the tests, you ship. The human is basically a "merge" button with opinions. What actually happens: the AI is allowed to accelerate the work and…
What everyone "knows": GitOps just means "deploy from Git instead of my laptop." And self-healing is great — it keeps the cluster matching what I want. What actually happens: the cluster now has a…
What everyone "knows": renting cloud GPUs is a money pit. Buy the hardware once and after that it's basically free — just the electricity. What actually happened: owning can be 6× cheaper… or it can…
What everyone "knows": self-hosting an LLM is easy. Rent a GPU, run vLLM, point your app at it. Done. What actually happens: the GPU you pick silently decides which models you can run, how fast, and…