I'm a former military officer who sells physical products for a living. No CS degree, no bootcamp. I taught myself to code because the tech industry wasn't going to let me in any other way.
GrandCru is a CLI that does real code review — complexity, naming, error handling, unused code, separation of concerns — delivered in the voice of a pretentious French wine sommelier. The character makes the feedback memorable and the output shareable. But take away the wine metaphors and you still get real issues with line numbers and fixes.
Why I built it: solo developers don't get code review. I don't have a team or a senior engineer. I needed something that would catch the stuff linters miss — and I wanted it to be fun enough that I'd actually run it.
Technical details:
- One API call per file using Anthropic's structured outputs (constrained decoding, not "please respond in JSON")
- Dual-channel Zod schema: strict technical data (line numbers, severity, fixes) and creative prose (tasting notes, remarks) in one response
- Extended thinking on by default — the model reasons about code architecture before committing to structured output
- BYOK: your Anthropic key, ~$0.02/file, no server
- TypeScript, Commander.js, npm
I pointed it at its own source code. It found real issues in its prompt builder — no input validation, raw string interpolation. Scored itself 79/100 — "Needs decanting before service."
This is a portfolio project from someone trying to prove they can build real tools. Happy to talk about the architecture, the self-teaching process, or structured output patterns.
I'm a former military officer who sells physical products for a living. No CS degree, no bootcamp. I taught myself to code because the tech industry wasn't going to let me in any other way.
GrandCru is a CLI that does real code review — complexity, naming, error handling, unused code, separation of concerns — delivered in the voice of a pretentious French wine sommelier. The character makes the feedback memorable and the output shareable. But take away the wine metaphors and you still get real issues with line numbers and fixes.
Why I built it: solo developers don't get code review. I don't have a team or a senior engineer. I needed something that would catch the stuff linters miss — and I wanted it to be fun enough that I'd actually run it.
Technical details:
- One API call per file using Anthropic's structured outputs (constrained decoding, not "please respond in JSON") - Dual-channel Zod schema: strict technical data (line numbers, severity, fixes) and creative prose (tasting notes, remarks) in one response - Extended thinking on by default — the model reasons about code architecture before committing to structured output - BYOK: your Anthropic key, ~$0.02/file, no server - TypeScript, Commander.js, npm
I pointed it at its own source code. It found real issues in its prompt builder — no input validation, raw string interpolation. Scored itself 79/100 — "Needs decanting before service."
This is a portfolio project from someone trying to prove they can build real tools. Happy to talk about the architecture, the self-teaching process, or structured output patterns.