NeuroBank Guardian

Redefining banking accessibility with voice-first Agentic AI. We replace complex UIs with natural conversation, empowering users of all abilities to manage wealth independently via Local LLMs.

Comment

💡 Inspiration: More Than Just “Accessible”

We started this hackathon with a simple question: Why is banking still so hard?

For most people, banking apps are annoying. But for millions — people who cannot see clearly, cannot hear, or face cognitive challenges — banking becomes a barrier. We saw friends and family struggle with tiny buttons, complex menus, and confusing screens. The current idea of “accessibility” often feels like a checklist, not a real solution.

We didn’t want to build another tool.
We wanted to build a Guardian — a system that feels human, reduces cognitive load, and helps people access their money with dignity and confidence.

🚀 What It Does

NeuroBank Guardian is an in-bank, accessibility-first AI assistant designed to help people who cannot see or cannot hear access banking services independently.

It provides:

  • Voice-first interaction for visually impaired users
  • Visual, text-based interaction for hearing-impaired users
  • A real-time AI avatar for natural communication
  • Proactive reminders for important bills and actions

Routine tasks like balance checks, bill reminders, and guided money transfers are handled automatically, reducing repetitive workload for bank staff.

🧠 How It Works: Agentic Workflow with Local AI

NeuroBank Guardian is not a traditional chatbot. It is built as an agentic system.

Using real-time WebSocket communication, the AI maintains a continuous two-way conversation with the user. When a user speaks or interacts visually, the system:

  1. Understands user intent in real time
  2. Checks relevant context (balance, bills, account state)
  3. Decides the safest next step
  4. Performs the action only after user confirmation
  5. Stores the interaction as part of the agent’s memory

This creates a natural, human-like banking experience instead of menu-based navigation.

🔐 Security & Privacy by Design

NeuroBank Guardian is designed to be installed locally inside bank infrastructure.

  • The AI runs on local LLM/SLM models, not public cloud chatbots
  • Sensitive banking data never leaves the bank
  • Information retrieval uses a vector database, not raw data access

All banking information is converted into vector embeddings (numerical representations). When a user asks a question, the system performs a Top-K similarity search on these vectors to retrieve relevant context — without exposing actual account numbers, documents, or personal data.

This ensures:

  • Strong privacy guarantees
  • No raw data exposure
  • Bank-grade security and compliance

🏦 Why It Matters

For customers:

  • Accessible banking without waiting in line
  • Independence for elderly and disabled users
  • Multilingual, easy-to-understand interaction

For banks:

  • Reduced staff workload
  • Faster in-branch support
  • Better accessibility without hiring additional staff

NeuroBank Guardian doesn’t replace human employees — it supports them.

🛠️ Built With

  • Python & FastAPI – backend and agent logic
  • WebSockets – real-time, continuous interaction
  • Vector Database (Top-K Search) – secure semantic retrieval
  • MongoDB Atlas – agent memory and interaction logs
  • Local AI Models (LLM/SLM) – private, on-prem intelligence
  • Next.js – frontend interface
  • Azure Avatar + STT/TTS – real-time voice and avatar interaction
  • Google AI Studio / OpenAI – AI orchestration and reasoning

Built With

  • azure-avatar
  • fastapi
  • gcp
  • google-ai-studios
  • next
  • open-ai
  • pinecone
  • python
  • stt
  • tts
  • vectordatabase
  • vectorembeddings
Share this project:

Updates