Eli Jah

I built CreativeFlow. Its functionality in simple terms: Speak a goal, get a structured action plan read back to you, confirm by voice, track progress by voice. Zero typing. Built entirely with Kiro's spec-driven, agentic workflow, and powered by ElevenLabs Conversational AI Agent for the full STT → LLM → TTS loop over WebRTC. The agent exposes three client tools that directly mutate a Zustand store to drive live React UI state from inside the voice session. A set of dynamic variables are injected at session start via a Clerk-authenticated token endpoint. Cross-session memory runs through ElevenLabs'. After each call, an AI-generated transcript summary is written to Redis, and then re-injected as a dynamic variable on the next session so the agent opens with context of what the user was last working on. What makes it special is that the ElevenLabs agent is configured to call tools that mutate live UI state and not just generate text. It uses a two-phase confirmation loop where it renders a draft task in the UI while ideating with the end user, and waits for spoken approval before persisting/saving it. Nothing is saved without intentional user confirmation.

Submitted 23 Apr 2026