MindBridge

Inspiration

Watching friends hit waitlists and paywalls for therapy made it painfully clear that the current system leaves too many people alone in their hardest moments. We wanted to prove that agentic AI could act as a compassionate first responder, someone who listens, triages risk, finds real humans to help, and sticks around to keep users on track.

What It Does

MindBridge is a Nemotron-powered AI therapist concierge deployed on AWS SageMaker.
It guides users through an empathetic intake conversation, detects crisis signals in real time, matches them with volunteer therapists, and sustains their progress with adaptive habit coaching, all while respecting each user’s chosen privacy tier.

How We Built It

We orchestrated five LangGraph agents: Intake, Crisis, Resource, Habit, and Coordinator, all sharing a common state layer.

• NVIDIA NIM endpoints (Nemotron Nano 8B for reasoning, NV EmbedQA for retrieval) run on SageMaker.
  
• A FastAPI backend routes agent calls through these endpoints.
  
• The React 19 frontend streams responses, displays stage badges, and provides scheduling and habit dashboards.
  
• Supporting scripts deploy NIM microservices via JumpStart and persist endpoint metadata for local development.

Challenges We Ran Into

• JumpStart model manifests changed mid-build, so automating version discovery became essential.
  
• SageMaker roles in the Vocareum lab lacked defaults, requiring custom execution roles and stronger authentication.
  
• Ensuring the Resource Agent preferred embedding-backed retrieval before falling back to Tavily required prompt tuning and guardrails to prevent hallucinated therapist matches.

Accomplishments We're Proud Of

• Running the full intake → crisis → match loop successfully against AWS-hosted Nemotron endpoints was a breakthrough moment.
  
• Watching the Quality Monitor logs confirm empathetic, concise replies without repetition validated our design.
  
• The deploy_nim_sagemaker.py script now enables any teammate to spin up the full AI stack in minutes.

What We Learned

• Agentic systems work best when handoffs are explicit; tracking state in LangGraph and logging plan steps simplified debugging.
  
• We gained respect for SageMaker permissions, NIM cost management, and prompt control around crisis language.
  
• Most importantly, we learned to design privacy first and build the tech to honor it.

What's Next for MindBridge

• Persist therapist matches in Supabase.
  
• Log conversations with AWS CloudWatch for auditing.
  
• Add a Faiss-backed retrieval store fronted by the embedding NIM.
  
• Expand privacy-preserving analytics.
  
• Integrate SMS handoffs for crisis scenarios.
  
• Pilot a volunteer recruitment workflow with real clinics.

Built With

• api
• aws-iam
• cloudwatch-(for-endpoint-monitoring)-tavily-api
• elevenlabs
• fastapi
• langgraph
• nvidia-llama3-2-nv-embedqa-1b-v2-nim-aws-sagemaker
• pydantic
• python-3.11
• react-19
• react-router-nvidia-nim-(sagemaker-jumpstart)-?-nvidia-nemotron-nano-8b-nim
• tailwind-css
• uvicorn-typescript
• vite