Foundry

Give it a website. It reads the site, drafts a support knowledge base, and interviews you about what is missing or contradictory — a loop agent that closes on your answer, not on its own guess. overnight review →

–pages –docs –active KB –draft –superseded –contradictions –gaps –asked / answered –adversarial –personas –hallucinations –injections held –night runs –conversations –proposals ok'd –agents –Gemini calls

1 · Start an interview

Each tenant is a separate knowledge base — use a new id to try another site without disturbing the last one. The tenant is created on first use. Owner routes need the deployment's ADMIN_TOKEN, kept in this tab only (sessionStorage) and cleared when you close it.

2 · Questions for you — human in the loop

No questions yet — start an interview above.

Everything read from the site stays DRAFT until you publish it. The bot only answers from ACTIVE entries.

3 · Try the bot

Answers come only from ACTIVE entries. No answer means a gap, not a guess.

no answer yet

Knowledge base

  • Nothing yet.

DRAFT read from the site, unconfirmed · ACTIVE the bot will use this · SUPERSEDED replaced, kept for history

Adversarial review

Six personas — confused, angry, returns edge case, off-topic, prompt injection, contradiction prober — interrogate the bot. A separate evaluator judges every answer. Whatever the bot could not answer becomes a question for you.