Concierge

What it is

An AI concierge grounded in this portfolio's canonical content model. It answers questions about the work and takes real actions: open a demo, jump to a case study or section, switch the theme. Ask by typing or voice, and have replies read back. If the model is unreachable or rate-limited it falls back to deterministic retrieval, so the chat never dead-ends.

• Gemini Flash
• Cloudflare Workers
• Action protocol
• Offline fallback

Under the hood

• Three renderers, one content model: the same canonical catalog feeds the interactive app, the crawlable text edition and the AI digest.
• Reply actions (open_app / open_study / scroll / set_theme / set_effects) are validated against the catalog before the shell executes them.
• The shared model key stays server-side behind a locked persona, token caps and a per-IP rate limit of 12 questions / 10 minutes.
• Deterministic retrieval fallback answers from the same catalog when the worker is unreachable, so every visitor gets a useful reply.

Honest notes

• Grounded but still a language model: ids and links are validated before execution, but prose can paraphrase imprecisely.
• Voice input uses the browser's speech service (Chrome sends audio to Google); it needs a network connection and permission and is hidden in unsupported browsers.