How a turn is put together, what every call is scored on, and the code you write to configure an agent. If you are evaluating Synapse technically, start here.
Four steps between the caller finishing and the agent answering. Every one of them is measured — and the part we control adds up to well under 300 ms.
Audio in over the phone or the browser, transcribed while the person is still speaking.
streamingA turn model, not a silence timer — so it does not cut people off mid-thought.
~90 msThe model streams a reply while the flow decides where the call goes next.
~140 ms to first tokenThe first clause goes to speech at once, and stops the moment the caller interrupts.
~50 ms to audioThe realtime path a conversation runs on, and what configuring each part actually looks like.
Pick up the phone and hold a real conversation. Fast replies, barge-in that stops mid-sentence, and a noisy line handled without falling apart.
# answer, listen, reply — inside a second def handle_info({:speech_final, text}, call) do {:ok, reply} = Synapse.LLM.stream(call.llm, text) Synapse.TTS.speak(call, reply, barge_in: true) {:noreply, call} end
Real numbers, in and out, at campaign volume. Inbound queues, outbound lists, and a warm transfer to a person when the agent has done its part.
Synapse.Campaign.run("hr-screen-july", rows: csv, concurrency: 400, on_qualified: &Synapse.Handoff.to_human/1 )
Every call is a trace: latency per hop, tokens, cost per minute, full transcript, and a score for how the agent actually behaved.
%Eval{ scorer: :interruption_gap, target_ms: 300, over: Synapse.Calls.today() } |> Synapse.Evals.run()
Four concepts, each backed by named metrics and a pass mark that lives in code. This is the report you get after every call.
How long does the caller wait before the agent starts talking?
A caller reads a long pause as the agent not having understood them, and starts repeating themselves. That is where calls fall apart.
Does the agent ever go quiet long enough to look dead?
This is the failure that loses calls. Background noise keeps the end-of-turn detector hearing words, so it never concludes the caller stopped, so it never answers. The caller says “hello?” and hangs up.
When the caller talks over the agent, does it stop and listen?
An agent that keeps talking over someone feels like it is not listening, and callers escalate fast once they think that.
Does the back and forth feel natural?
Long handovers and constant talking over each other both make a call tiring, even when every answer is correct.
One call, scored. A long red block would be the agent going quiet after the caller finished — the thing a caller reads as a dropped line. This call has none.
The thresholds live in code, not in a settings page — so nobody can quietly widen one to make a bad week look green.
We are preparing to open the runtime underneath Synapse — the media path, the turn detection and the supervision tree that keeps one bad call from touching any other.
The realtime core: media in and out, turn detection, the supervision tree, and the transport adapters. Enough to run an agent end to end on your own machines.
The agent builder, the eval reports, campaign orchestration and the dashboards. That is the product, and it is how we keep the lights on.
Because a repository nobody can run is not open source. We would rather ship it once the setup path is honest than publish a snapshot and leave it.
We will run one of your real workflows and send you the report.