RecipeWalk

Challenge

Cooking from a phone is hands-busy work, and recipe blogs bury each step under ads, life stories, and walls of text. RecipeWalk structures the recipe, then a real-time voice assistant guides the cook hands-free, without “ChatGPT forgetfulness” (losing track of the current step).

Process

Gemini 2.5 Flash parses a pasted or photographed recipe into ingredients, steps, and timers. Voice connects client-direct to the Gemini Live API over WebSocket, skipping a backend proxy to cut latency and cost. A server-side state machine sends step-scoped system instructions and issues single-use, scoped, ephemeral session tokens, so the assistant always knows the current step.

Other projects copy RecipeWalk’s reference implementations: secrets loaded from AWS at service start into a locked-down env file, IAM deployer scoping, and gift-code security (high-entropy codes encrypted at rest, atomic single-use redemption, IDOR hardening, auto-revocation on refund or dispute). My voice-eval harness runs synthesized audio through the exact production voice config to catch tool-calling regressions; it caught one before release.

Result

Deployed to production at recipewalk.com, all tests green (277 backend, 120 frontend). A real-money buy, redeem, refund, and revoke cycle proved the gift flow end-to-end. Not yet open to customers; a paid beta waits only on a go-live decision I am holding on purpose.