All projects
Palette
An AI agent that generates complete short dramas end-to-end.
- Type
- Short-drama generation agent
- Context
- Stealth Startup
- Year
- 2026
Highlights
- Architected the runtime around a single “producer” orchestrator dispatching 13 specialized sub-agents, with skills as version-controlled prompts and a per-agent least-privilege tool matrix; the orchestrator owns dispatch only, preventing runaway recursion.
- Built an in-process MCP tool layer with typed contracts unifying text, image, and video generation across fal.ai, Kling, and Gemini, offloading bulk generation to fast models for large latency and cost wins.
- Designed a three-layer sandbox for multi-tenant execution (per-session storage, OS-level filesystem/network isolation, runtime authorization hooks) that blocks prompt injection and path traversal independently at each layer.
- Engineered a lazy write-through S3 file cache, SSE streaming with session resume, an asynchronous video-job pipeline that keeps minutes-long renders off the agent loop, and atomic per-tool usage metering as the SaaS billing guardrail.