Hey HN,Built VoooAI over 4 months. It's a NL2Workflow engine – you describe what you want in plain English, it generates an executable multi-modal pipeline (image + video + audio).The problem:
Midjourney for images, Runway/Pika for videos, Suno for music – each has its own API, interface, and learning curve
ComfyUI gives you control but requires understanding node graphs
n8n connects APIs but doesn't understand creative intent
What we did differently:Instead of having LLMs output JSON directly (hard to validate), we introduced a pseudocode IR layer:
LLM generates pseudocode → deterministic interpreter expands to executable workflow JSON. This gives us:
Syntax validation before interpretation
Automatic engine correction (if skill requires specific engine)
Reliable edge connections
Try it: voooai.com Compare us: voooai.com/workflow-comparison – we're not better at everything, but for "idea → multi-modal output" in one prompt, that's our niche.AMA about the architecture.
Comments
hkljjkl•1h ago
What Vibe Coding Taught Me
Code review skills matter more than coding skills: I spend 70% of my time reviewing generated code, understanding it, catching subtle bugs. The AI writes fast but confidently wrong.
Architecture conversations are gold: The most valuable prompts aren't "write this function" but "let's discuss the tradeoffs between X and Y architecture."
Technical debt accumulates differently: I have less "I wrote this at 2am" debt, more "the AI suggested this pattern and I didn't push back" debt.
Debugging is harder: When you didn't write the code, you don't have muscle memory for where things might break.
Current State
Not open source: The system integrates with 30+ APIs requiring paid subscriptions. Maintaining compatibility is a full-time job I can't delegate to the community.
Business model: Freemium. 50 free points/month (10-20 images or 2-3 videos). Subscriptions for power users.
What's working: NL2Workflow accuracy (~85%), multi-modal pipelines (image→video→audio), streaming UX
What's not: Complex requests (>50 nodes), specific style transfer edge cases, real-time collaboration
The One-Person Team Reality 4 months, one person, vibe coding. But let me be honest about what that means:
I work 10-12 hour days
The AI doesn't replace a team – it amplifies one person
I still need to understand every subsystem deeply
Customer support, marketing, DevOps – I do those manually
Vibe coding let me compete on product scope, but it didn't eliminate the other hats a founder wears.
hkljjkl•1h ago