Rocdel

An agentic AI revenue platform. Six agents enrich, score, draft, call, meet, and forecast, grounded in RAG over the team's own revenue history. Every claim carries a record citation and nothing reaches a prospect without a human approving it. Tools are reached through MCP servers with scoped credentials; an eval harness gates every prompt or model change before deploy.

The architecture is five layers with one rule between them: a human approval gate sits between every agent output and any external send. Agents reach tools through MCP servers that hold scoped credentials, so an outreach agent can send as a rep but never read the inbox.

Postgres with pgvector holds the revenue corpus, Redis Streams runs the task queue, and OpenTelemetry traces every step from retrieval to tool call. Prompt and model changes replay the eval harness before they ship; a regression on any metric blocks the deploy.