voice agent for stocks
LLM-powered voice agent for daily stock-market chat, with offline KB pre-processing
A voice agent for daily stock-market conversation — earnings, macro, ticker news. The design sidesteps the usual latency-vs-reasoning trade-off in voice products: every morning a frontier model pre-processes the day’s news, analyst notes, and price action into a semantically-indexed brief; at runtime the voice layer only retrieves from that brief and reads it out. You get frontier-quality content with small-model latency, at the cost of being scoped to a single domain.
Stack: FastAPI + WebRTC, webrtcvad, Whisper ASR (SenseVoice as self-hosted fallback), Edge-TTS, GLM-4.5-air, Supabase + Upstash Redis.