SahaayaAI — AI That Knows When to Ask for Help
Link to open source: https://github.com/Rajakumari208/sahayaai
SahaayaAI is a multilingual, real-time voice AI assistance agent designed for difficult customer-support and public-information conversations.
Real callers may switch between languages, speak through background noise, interrupt the agent, or provide incomplete information. In these situations, simply generating a confident answer is not enough — the AI must know when it is uncertain.
SahaayaAI follows a five-stage operating loop: Understand → Collect → Confirm → Assess → Escalate.
It identifies the caller's intent and essential details such as location and service identifiers, while maintaining field-level confidence. When a critical detail is unclear, SahaayaAI does not guess. It asks for confirmation and, if uncertainty remains, escalates the conversation to a human with the relevant transcript, confirmed information, missing information, confidence indicators, and reason for escalation.
The prototype demonstrates Hindi + English code-switching, noisy-call handling, interruption awareness, confidence-aware decision making, critical-detail confirmation, and human-in-the-loop escalation using real-time voice communication.
Our core principle is simple:
AI handles the conversation. Humans handle the decisions.
SahaayaAI is designed not to replace human support agents, but to handle the conversation safely until human judgment is needed.
This build was uploaded as a hackathon project
