Sakshi AI Companion
Live demoSub-second voice replies with real long-term memory
A voice-first AI companion with real-time streaming conversation, long-term memory via mem0, and built-in tools for web search, calculation, and document RAG.
The problem
Most voice AI demos feel laggy and forgetful — noticeable turn-taking delay breaks the illusion of a conversation, and there's no memory once the session ends.
The approach
Built a real-time streaming pipeline on Groq for fast inference and Deepgram for speech, so replies land in well under a second. Long-term memory runs through mem0, so the assistant carries context across sessions instead of starting cold every time. On top of the voice loop sit built-in tools — web search, calculation, and document RAG — so the assistant can act on a request, not just talk about it.
The result
A voice companion with sub-second replies and memory that actually persists between conversations — covered in more depth, including the honest read on what's sellable underneath the persona, in the full write-up.
Stack
React
Frontend UI
FastAPI
Backend orchestration
Groq
Low-latency LLM inference
Deepgram
Speech-to-text and text-to-speech
mem0
Long-term memory layer