Voice & Calls - ScaleDown

Calls, voice agents and meeting recordings, after transcription. Four recipes, one per ScaleDown model.

Compress

Trim the backend docs down to what the question needs, so the agent can answer a live caller fast.

Sits in front of your answering LLM. Open the full Python recipe.


Why it fits	The caller is not left waiting while a big chunk of retrieved text gets processed.
How it goes	Caller asks → speech transcribed → backend pulls docs → Compress trims them → LLM writes a short answer → agent says it.
Payoff	Lower latency on the retrieval step, less token spend per answered call.

Summarize

Turn a call or meeting transcript into a short recap with the decisions and action items.

Replaces an LLM call. Open the full Python recipe.


Why it fits	Transcripts are long, conversational, and multi-speaker, exactly the workload Summarize is built for.
How it goes	Call ends → transcript in → Summarize produces the recap → push to the notetaker, CRM, or follow-up draft.
Payoff	State-of-the-art on QMSum, the meeting-summarization benchmark.

Extract

Pull the CRM fields out of a sales call, deal size, economic buyer, next meeting, into clean fields.

Replaces an LLM call. Open the full Python recipe.


Why it fits	A deal cannot move on a guess; Extract returns the numbers and names exactly as they were said on the call.
How it goes	Call ends → transcript in → you tell Extract what to look for → it returns the fields → push into HubSpot.
Payoff	Structured CRM fields straight off the transcript, no rep data entry.

Classify

Send the caller to the right place based on what they say, instead of press 1 for billing.

Replaces an LLM call. Open the full Python recipe.


Why it fits	Routing is the busiest step in the call; Classify handles a simple choice fast without paying a frontier model to overthink it.
How it goes	Caller says what they need → speech transcribed → Classify picks the label → agent routes them into that flow.
Payoff	Faster routing at lower cost per call.