NDZ Labs: an independent AI R&D studio in Buenos Aires
Hi, I'm Francisco Fernández, founder of NDZ Labs, an independent AI R&D studio in Buenos Aires. I work as an AI solutions architect and product manager, experimenting with agentic systems for product development, research, and analysis.
My main project is webeloper, a multi-agent pipeline that researches, designs, and builds client websites. Each role works from an explicit contract of requirements, boundaries, and validation criteria. I build with Claude, Gemini, Qwen, and DeepSeek, deploy on Vercel, and store data in Supabase. What I'm still figuring out is how much autonomy an agent should have before a human steps in, and when a specialized team beats a single agent.
I'd love feedback from anyone who has built multi-agent pipelines, especially on how you test whether handoff contracts actually hold up.






Comments2
When webeloper builds a website for a client, do those clients want a real person to review and approve it before it goes live?
The site absolutely requires the client’s review and approval. In fact, as a business decision, it is the client who deploys the site to their hosting provider (unless they specifically request that the developer handle the deployment). Therefore, while the site is designed and built by agents, all judgment-based decisions remain 100% human. Webeloper is designed for lightweight websites; it primarily generates "vanilla" (plain) code, making it easy to read should the client wish to review the source code.
2 comments
When webeloper builds a website for a client, do those clients want a real person to review and approve it before it goes live?
The site absolutely requires the client’s review and approval. In fact, as a business decision, it is the client who deploys the site to their hosting provider (unless they specifically request that the developer handle the deployment). Therefore, while the site is designed and built by agents, all judgment-based decisions remain 100% human. Webeloper is designed for lightweight websites; it primarily generates "vanilla" (plain) code, making it easy to read should the client wish to review the source code.