Yes, the idea is to ‘speak/write’ to the local model to fix those little things so you don’t have to do them by hand. I actually already have a fine-tuned Qwen model running on Apple’s MLX to handle some of that, but given the hard YC deadline, it didn’t make it into the demo.
Eventually, you’d say, ‘add an additional layer, TopicsController, between those two files,’ and the local model would do it quickly without a problem, since it doesn’t involve complicated code generation. You’d only use powerful remote models at the end.
Since this topic is closely related to my new project, I’d love to hear your opinion on it.
I’m thinking of building an AI IDE that helps engineers write production quality code quickly when working with AI. The core idea is to introduce a new kind of collaboration workflow.
You start with the same kind of prompt, like “I want to build this feature...”, but instead of the model making changes right away, it proposes an architecture for what it plans to do, shown from a bird’s-eye view in the 2D canvas.
You collaborate with the AI on this architecture to ensure everything is built the way you want. You’re setting up data flows, structure, and validation checks. Once you’re satisfied with the design, you hit play, and the model writes the code.
I've spent a lot of time building something I wish I'd had when I was learning Spanish: an app that lets intermediate learners practice with real, interesting content. When I studied Spanish, I often got frustrated because native content had too many words and expressions I didn't understand.
Thanks to recent AI advancements, it's finally doable—but getting it right was tricky. It took setting up a multi-step pipeline and fine-tuning several AI models to make it work smoothly.
Here's how it looks in practice, turning a Spanish news article into a personalized language lesson:
- Simplify the article to match the student's current skill level.
- Add language elements the student is currently studying (like the past tense or certain phrases).
- Include specific vocabulary words the student is learning to reinforce them in context.
- Use another AI model to tokenize and lemmatize the text so the app can accurately recognize each word's context and offer quizzes, definitions, and comprehension checks.
- Automatically link words and expressions to detailed dictionary entries.
- Generate high-quality audio with ElevenLabs for authentic listening practice.
There are a few smaller steps involved too, but those are the main ones.
I think Topics is the first app to create fully personalized, contextual language lessons directly from current news content.
I'd love to hear your thoughts! If you're interested in trying it out, send me a PM—I can give you an access code in exchange for your feedback :)
PS. I’m stealing the ‘antidote to “vibe coding”’ phrase :)