Sage

What it solves

AI assistants today live behind text boxes. You type a prompt, wait for output, then copy it somewhere useful. That interaction model breaks down in the most common knowledge-work setting: the meeting. Sage removes the friction by joining calls as a voice participant — no screen-sharing, no app-switching, no typing. Someone says "Sage, what's the latest on that?" and Sage answers out loud, in context, like any other person on the call.

This is a proof of concept for what AI employees could look like when they show up the same way human employees do: with a voice, a name, and the ability to participate in real-time conversation.

How it works

Sage bridges a live phone call with OpenAI's Realtime API using Twilio for telephony:

1. Sage receives a call and announces it's present when asked to unmute
2. Audio streams in real time — Twilio captures the call audio and streams it over WebSocket to the server
3. The server bridges audio to OpenAI — incoming speech is piped to the Realtime API, which handles voice activity detection, transcription, and response generation
4. Sage only speaks when addressed by name — it listens to the full conversation but stays silent unless someone says "Sage." This prevents over-talking and keeps the meeting natural
5. Tool calls happen mid-conversation — when asked to look something up, Sage says "Let me check on that," runs a web search, and summarizes findings conversationally
6. Recap emails on request — ask Sage to email a summary and it generates a formatted recap and sends it via Resend to whatever addresses you specify
7. Live dashboard — a WebSocket-powered dashboard shows active calls, real-time transcripts, agent status (listening, thinking, speaking), and tool call activity

The entire pipeline runs with sub-second latency using g711 audio encoding and server-side voice activity detection, so Sage's responses feel conversational rather than robotic.

What makes it different

Most "AI meeting assistants" record, transcribe, and summarize after the fact. Sage participates during the meeting. It's the difference between taking notes and having a colleague in the room who can answer questions, pull data, and contribute to the discussion as it happens.

Key design decisions that make this work:

• Name-gated responses — Sage only speaks when directly addressed, so it never interrupts or derails the conversation
• Concise by design — responses are kept brief because long monologues waste everyone's time in a live meeting
• Real-time web search — Sage can look up current information mid-conversation, turning "let's table that and look it up later" into "Sage, can you check that right now?"

Future state

Sage today is a working proof of concept. The roadmap includes capabilities that move it from meeting participant to persistent team member:

• Tool calling — trigger workflows, update project trackers, create tickets, or pull data from internal systems during the call
• Meeting scheduling — Sage could schedule follow-up meetings, send calendar invites, and coordinate availability
• Persistent memory — store context across meetings so Sage remembers past discussions, decisions, action items, and participant preferences
• Multi-meeting awareness — connect insights across meetings to surface patterns, track commitments, and flag when prior decisions are being revisited

Ideal for

• Enterprise leaders evaluating AI who want to see what a voice-first AI employee looks like in practice — not slides about it
• Teams running frequent meetings who lose time to "let me look that up and get back to you" moments that Sage can resolve in real time
• AI strategy and innovation teams exploring how voice agents could integrate into daily workflows beyond chatbots and copilots
• Anyone building the case for AI adoption who needs a tangible, memorable demo that lands with non-technical stakeholders

Try it out

View the live dashboard