AI Content Engine with a Human Approval Gate

The situation

The venture is a publisher, which means an analyst must approve every public word. The plan also called for high-volume daily content from two founders. The only way both are true is a pipeline that does the drafting and a gate that makes publishing without approval structurally impossible.

What I built

One always-on Python container runs the Bot Framework handler and every scheduler as an asyncio task. A poller watches public social feeds every six minutes and writes candidates to the warehouse. Claude rewrites each one into an original story under the house-voice and compliance rules, with hard guards on length and style. A chart renderer adds a branded, watermarked image.

The draft and chart go to Microsoft Teams as an Adaptive Card. Approvers are allowlisted, and the Approve action fires an invoke that publishes to X and Discord, records who approved, and stamps the performance tracker with the live market price. The bot authenticates with a managed identity, sits behind its own edge endpoint and firewall, and carries an origin fence with tests. Source attribution never reaches members.

The outcome

Daily content ships with one analyst click per post, and the performance archive stays honest because the publish and the price stamp are the same event.

Stack