AgentX

AI & systems / 2026

AgentX

AI Hallucination Juror combines a content-review workflow with Command Shield. It exposes verdicts through a FastAPI backend, MCP tools, a terminal interface, and editor and browser integrations.

My contribution

Worked on a local verification system connecting technical-content review, command analysis, and verdict history across developer interfaces.

The approach

• Request five review perspectives in one batched Gemini call, with a correction pass for blocked content.
• Inspect commands with patterns, package-registry and OSV lookups, and model reasoning.
• Store review history in SQLite and surface results through CLI, extensions, HTTP, and MCP.

Scope & perspective

Review adds evidence and friction before trusting generated output. Model-assisted verdicts remain fallible; five perspectives in one call are not five independent models.