-
-
admin panel
-
admin panel
Inspiration
Every person alive is being controlled by beliefs they never chose. A harsh manager in 2015 who fired someone for asking questions. A parent who said showing emotion is weakness. A school that taught moving fast means being sloppy. These moments become invisible rules — and nobody has ever built a tool to find them.
That is what inspired MindRoots.
What it does
MindRoots is an AI belief archaeologist. It conducts a live voice interview using the Gemini Live API, asks layered Socratic questions until it hits the bedrock of your belief system, and then maps every belief back to its exact origin.
The output is a personal Belief Origin Document containing:
- A written analysis of each belief and where it came from
- AI-generated illustrations of the specific memories uncovered
- A synthesized voice narration summarizing the entire journey
All delivered in one seamless multimodal flow: voice → reasoning → image → audio.
How we built it
MindRoots uses a three-agent ADK pipeline hosted on Google Cloud:
Agent 1 — The Socratic Interviewer: Powered by Gemini Live API, this agent conducts a real-time voice conversation. It asks layered "why" questions, digging deeper with each answer until it reaches the original moment or person that planted each belief.
Agent 2 — The Data Structurer: Takes the raw conversation transcript and converts it into a structured JSON belief graph stored in Cloud SQL. Each node captures the belief, its origin, who planted it, its age, and whether it still serves the user.
Agent 3 — The Storyteller: Reads the belief graph and generates the Belief Origin Document — combining written analysis, Imagen-generated illustrations of specific memories, and an ElevenLabs voice narration of the full journey.
The backend is deployed on Google Cloud Run. Sensitive session data is encrypted and stored in Cloud SQL. The frontend is built in React.
Challenges we ran into
- Designing the Socratic questioning logic so the agent knows when it has reached the true "bedrock" belief and stops drilling
- Maintaining context across the three-agent handoff without losing emotional nuance from the original conversation
- Generating AI illustrations that accurately reflect the specific memory described, not generic imagery
- Keeping the Gemini Live API voice conversation feeling natural and human despite the deep psychological subject matter
Accomplishments that we're proud of
- The first AI tool in the world designed specifically to excavate limiting beliefs
- A fully seamless voice → map → document → audio pipeline with zero text-in/text-out
- Demo reactions — people who tested it were genuinely moved seeing their own mental history mapped on screen
What we learned
- Gemini Live API handles interruptions and natural speech far better than we expected
- The ADK pipeline architecture made multi-agent coordination clean and reliable
- Multimodal output (voice + image + text together) creates emotional impact that text alone cannot achieve
What's next for MindRoots
- Therapist and coach dashboard to track client belief maps over time
- Team version for organizations to identify shared limiting beliefs across departments
- Longitudinal tracking — resurface old beliefs after 30/60/90 days to measure change
- Integration with calendar and journaling tools to spot belief patterns in daily decisions
This project was built for the Gemini Live Agent Challenge. #GeminiLiveAgentChallenge
Built With
- elevenlabs-tts
- fastapi
- gemini-live-api
- google-adk-(agent-development-kit)
- google-cloud-run
- google-cloud-sql
- imagen-(ai-image-generation)
- python
- react
- vertex
Log in or sign up for Devpost to join the conversation.