Open Vectorizer, a raster-to-SVG engine written in Rust that compiles to WebAssembly and runs locally. It's a technical project with a reproducible benchmark suite that actually holds its own against tools like Potrace and VTracer. But when I tried to share it in traditional dev circles, I hit a wall.
Some communities are now treating "AI-generated" as a red flag for "low quality." For instance, some subreddits now require you to certify that a project doesn't contain significant AI-generated content. If you do, it's dismissed as low-effort.
The problem is that "Was AI used?" is a terrible proxy for "Is this actually good software?"
The Spectrum of AI Implementation #
There is a massive gap between someone who prompts "make me a vectorizer" and publishes the raw output, and a developer using an AI workflow to iterate on complex logic. Consider these different levels of AI integration:
Level 1: Basic IDE autocomplete.Level 2: Using Copilot for boilerplate and completions.Level 3: Using agents to implement specific functions based on strict technical specs.Level 4: Designing the architecture and tests, then delegating the implementation to an LLM agent.Level 5: Zero-thought prompting and immediate publishing.
The industry is currently trying to filter out Level 5 by banning everything from Level 1 to 4.
Real-world Iteration with AI #
Open Vectorizer is a perfect example of why this distinction matters. The first version I built was mediocre—it used a standard pipeline of finding pixel boundaries and fitting curves. The circles weren't circular, and the geometry was messy. I eventually shelved it.
When I came back to it, I used a heavy AI workflow. We didn't just "generate code"; we brainstormed whether to use a vision model trained on synthetic SVG-to-PNG data. We weighed the pros and cons and ultimately decided against ML in favor of redesigning the deterministic algorithm.
The result is a tool that actually works and outperforms established benchmarks, yet because of the process used to get there, it's viewed as "AI-generated" by some purists.
For those of us building a modern AI workflow, the goal isn't to avoid the tool, but to use it to reach a level of quality that was previously impossible for a solo dev. If the benchmarks are reproducible and the code is MIT-licensed, the origin of the characters on the screen shouldn't be the primary metric for value.
[Next Inertia. →](/en/threads/3871/)