Our approach is an "AI-Managed" abstraction layer. The key architectural decision is that clients and developers do not interact directly. Also, most talents are senior software developers from China that are cheap but high-value.
It's an event-driven workflow:
- A client submits a project goal in natural language.
- Our AI PM (built using LLMs) deconstructs the goal into concrete technical tasks, milestones, and deliverables.
- These tasks are routed to our vetted developer pool.
- The AI PM manages the entire lifecycle: progress tracking, check-ins, and validating deliverables before packaging the final product for the client.
The main technical hurdle we're tackling now is agent reliability. Getting it to correctly parse complex project info, monitor progress accurately, and trigger notifications via a VM-based system is non-trivial. The prompt engineering required for this kind of long-running, stateful project management is where most of our effort is.
Our stack: Next.js, Sass, shadcn/ui, and gpt-4.1-nano for the agent.
We're in the very early stages and would love to get your feedback on the core concept and our technical approach.