EchoAgent
An AI agent that facilitates meetings instead of just recording them
An AI agent that joins your team's calls, listens as people speak, and talks back — running agendas, answering questions live, and producing summaries with action items.
- <2s
- Target time from question to spoken answer
- ~2×
- Faster replies in always-on mode
- 0
- Typing required — it all happens by voice
- 17
- Settings to tune voice, pacing, and behavior
EchoAgent BOT
Speaking…
Alex
Jordan
Sam
Agenda
- 01 · Standup recap
- 02 · API versioning
- 03 · Sprint commitments
Live transcript
Alex: What did we decide on API versioning last week?
EchoAgent: You agreed on URL-based versioning — v2 ships behind a feature flag this sprint. Want me to log a follow-up action item for the migration guide?
Jordan: Yes, assign it to me.
✓ Action item captured: “Draft v2 migration guide” — Jordan
STT 184ms · LLM 232ms · TTS 308ms · total 0.72s
Illustrative mockup — representative of the product interface, with placeholder data.
Overview
Most meeting AI sits silently in the corner and emails you a transcript afterwards. EchoAgent takes the opposite approach: it participates. It joins your team's voice call, listens as people speak, works out what's being asked, and replies out loud in a natural voice — fast enough to hold a real conversation.
The bot acts as an expert facilitator: it can run a prepared agenda, keep time, capture decisions and action items as they happen, and answer questions mid-meeting without anyone touching a keyboard.
Tech Stack
AI Angle
End-to-end conversational AI: live transcription, agent-based reasoning with memory, and natural voice synthesis composed into a single real-time loop.
The Challenge
Real-time voice AI is an engineering problem before it is an AI problem. Sound has to travel from the call, through transcription, reasoning, and speech generation, and back into the conversation quickly enough that the agent feels present rather than laggy.
The Approach
EchoAgent splits the work in two: a lightweight audio layer moves sound in and out of the call, while a dedicated AI brain owns everything intelligent — memory, the agenda, and the meeting summary. Each half can improve without breaking the other.
Speech is transcribed live as people talk (via Deepgram), and replies come back as natural speech (via ElevenLabs), with the whole round trip tuned to feel conversational. The AI brain is swappable — it can run on the same machine or on a remote server, and the always-on remote mode roughly halves response times.
The pipeline is built to fail gracefully: dropped connections quietly restore themselves, the agent filters out its own voice so it never responds to itself, and a simple turn-taking model — idle, listening, thinking, speaking — keeps the conversation orderly even when people interrupt.
The Speech-to-Speech Loop
- 01Voice CallIt hears each speaker in the call
- 02Live TranscriptionSpeech becomes text as people talk
- 03AI ReasoningContext, memory, and agenda state
- 04Voice SynthesisThe reply becomes natural speech
- 05Voice ReplySpoken back into the channel
Highlights
- A complete voice loop: it hears speech, thinks, and talks back — no typing anywhere
- Interrupt it mid-sentence and it stops talking and listens
- Speaks a brief filler phrase whenever an answer needs a moment — no awkward silence
- Always-on connection mode roughly halves response times
- Every response is timed stage by stage, so slowdowns are easy to find and fix
- Simple chat commands to invite it in, load an agenda, or request the summary
Want something like this built?
We design and build AI products, analytics platforms, and automation systems end to end. Tell us about your problem.