Blog
Notes from building Adrian
How the pipeline actually works, what shipped, and the parts we had to redo.
Do you need a GPU to run a local coding model?
You can run a local coding model with no GPU at all. What a GPU actually buys you, what VRAM changes, and when CPU-only stops being tolerable.
Who owns code that an AI wrote?
Who owns code an AI wrote? What Adrian's terms say, what changes when a hosted provider is in the loop, and the parts the law has not settled yet.
Why your local LLM is slow, and what to check first
A local model that crawls is usually swapping, not underpowered. The four things that actually cost you speed, in the order worth checking them.
Build an app without coding: what that really means
You can get a working app from a description today. What you cannot get is a finished product without ever reading code. Here is where the line actually sits.
Prototype vs MVP: the difference that actually matters
A prototype answers a question and gets thrown away. An MVP is the smallest thing real users depend on. Confusing them is how teams ship the wrong artefact.
AI code generation: what it does well and where it stops
A practical account of AI code generation — the tasks it genuinely handles, the ones it reliably fumbles, and how to tell which you have before shipping it.
Natural language to code: where the meaning gets lost
Turning a description into code fails at the same place every time: the parts you did not say. What ambiguity costs, and how to write a prompt that lands.
AI code refactoring: changing code that already works
Refactoring is harder for a model than writing new code, because the bar is invisible: everything must keep working. What that changes about how you ask.
Reproducible builds when the code was generated
The same prompt does not give the same code twice. Why that is fine, which reproducibility actually matters, and how to verify a build you did not compile.
Ollama vs LM Studio for local coding models
Both run the same open models on your own hardware. The difference that decides it is whether you are the one using it, or another program is.
Local LLM development: how the workflow changes
Developing with a local model is not a worse version of the hosted workflow. It is a different one: smaller steps, tighter scope, and structure doing the work.
How to validate an app idea by building it
A mockup tests whether people like a picture. A working app tests whether they use it. How to validate an app idea by building the real thing cheaply.
Secure app development when AI writes the code
AI writes code that compiles and quietly ships an eval() or a postinstall hook. What actually goes wrong, and which parts a build system can catch for you.
Which local coding model actually fits your machine
Adrian ships a curated catalog of local models for Ollama. Here is what genuinely runs on 8GB, 16GB, 32GB and beyond — and what you give up further down.
Bring your own API key: what building with Adrian actually costs
Three ways to pay for AI-built software — local models, your own API key, or hosted credits — and an honest account of what each one really costs you.
Run an AI app builder on your own machine
What it takes to build working apps with a local LLM: no API key, no per-token bill, nothing sent to a model provider — and the trade-offs that come with it.
Why AI-generated apps compile but do not work
Compiling is a low bar. How generated apps fail while every signal says success — leftover scaffold text, missing features, controls nothing can change.
How Adrian builds an app, step by step
What happens between typing a prompt and getting a running app: planning, specialist agents, code generation, a compile gate, and a live preview.