Voice AI Companion
A voice-first AI companion with real memory and sub-second replies.
Sakshi is a voice AI companion with long-term memory, web search, and document RAG — built for real-time, low-latency conversation, not chatbot round-trips.
<900ms
End-to-end voice latency
5
Built-in tools
Long-term
Memory via mem0
What it does
Real-time voice chat
Bidirectional streaming voice conversation powered by Deepgram STT/TTS and a Groq-hosted Llama 3.1 brain.
Actually remembers you
mem0-backed long-term memory means conversations build on each other instead of resetting every session.
Tool calling
Web search, calculator, weather, and document RAG are wired in as callable tools the agent reaches for automatically.
State-driven UI
A React + Tailwind interface with Lottie animations reflects idle, listening, thinking, and speaking states live.
Voice preview
Sakshi — calm, low-latency voice
Mockup of the in-call state UI — not a live demo
How it's built
Curious how Sakshi AI Companion works under the hood?
Full source, architecture docs, and setup instructions are on GitHub.
This is a personal project and isn't publicly hosted, to keep rate limits and data resourcing manageable. Want to try it? Join the waitlist above or fork the repo and run it yourself.