/// case_study · custom_dev · ai

A desktop co-pilot that listens to your sales call, and feeds you the answer, live.

We built a native desktop app that hears a live sales call, transcribes it on-device, and streams real-time talking points and follow-up questions drawn from your own data: a template any sales team can point at their product, pricing, and playbooks.
/built_by
Nexxion, in-house
/type
Native desktop application
/ai
Claude: Anthropic API or Claude on AWS
/knowledge
Any data, via tools (MCP)
/// the_problem

No salesperson can hold the whole catalogue, the pricing, and every playbook in their head mid-call. The perfect answer usually arrives five minutes after the call ends.

  • The right answer comes too late: you remember the ideal recommendation or the sharp follow-up question after the moment has passed.
  • Chatbots don't help live: you can't type into an AI chat window while you're talking to a customer.
  • Call audio is sensitive: shipping a customer's voice to a cloud transcription service is a non-starter for many teams.
  • Off-the-shelf is English-first: our calls are in Greek with heavy English technical code-switching, which most tools mangle.

We wanted a co-pilot that listens, understands, and hands the rep the right thing to say, while the customer is still talking.

/// how_it_works
01

Listen

Captures the customer's voice through system audio (Meet, Teams, Zoom, or a phone bridge), no browser extension, no bot in the meeting.

02

Transcribe on-device

Local speech-to-text on the laptop's GPU. The audio never leaves the machine; only text goes onward. Handles Greek/English code-switching.

03

Reason

Every ~20 seconds, Claude reads the rolling transcript and pulls the relevant facts from your data via tool calls: recommendations plus questions to ask.

04

Surface

Glanceable cards stream into a small native window beside the call, paced so the rep is helped, not flooded. The call ends with an instant summary.

/// by_the_numbers
~20s
live recommendation cadence: talking points while you're still talking
on-device
speech-to-text: the call audio never leaves the laptop
$1–2/hr
Claude Haiku, vs the ~$50–150/hr cloud-STT stack we rejected
2
model paths: the Anthropic API or Claude on AWS

// figures from the delivered build; LLM cost is at the live recommendation cadence on Claude Haiku.

“A co-pilot that listens to the call and hands the rep the right answer, from their own data, while the customer is still talking.”
/// what_we_built
01

It listens to the call, privately

on-device · source-agnostic

The app captures the customer's voice straight from system audio, so it works with any meeting tool or a phone bridge, with no browser extension or meeting bot. Speech-to-text runs locally on the laptop's GPU: the audio never leaves the machine, only the text moves on. It's tuned for Greek with heavy English technical code-switching, so terms like "SQL Server" survive transcription intact.

02

Live talking points & questions

~20s cadence

On a steady cadence, Claude reads the rolling transcript and surfaces two things: recommendations the rep can raise now, and follow-up questions to draw out more detail. They render as glanceable cards in a native window next to the call (paced so the rep gets help, not noise), and the session ends with a one-click summary.

03

Grounded in your data: the template

any catalog · via MCP

The model doesn't guess from a catalogue baked into a prompt; it pulls facts on demand via tool calls (MCP) from whatever knowledge you point it at: product docs, pricing, battle cards, case studies. We first wired it to live AWS documentation; swap the tools and the same app sells anything, always from fresh, citable sources.

04

Claude, or Claude on AWS

your billing, your choice

Pick the provider in settings: the Anthropic API (paste a key, go) or Claude on AWS: Claude Platform on AWS (generally available in 2026), which runs the same models and features through your AWS account for consolidated billing and EU residency, authenticated with an AWS-issued key and workspace. It also runs as an MCP server, so Claude Desktop can drive it for post-call deep-dives.

/// the_engineering

The stack is the product. We got to a hybrid local-STT + cloud-LLM design by rejecting the obvious ones.

What we rejected: a fully-cloud stack (cloud transcription + a frontier model on a 5-second cadence) ran ~$50–150/hour per meeting and sent audio off-device; a fully-local stack (a small open model on the laptop) was too slow: recommendations landed after the moment had passed.

What we shipped: transcription stays local on the GPU (private and real-time), and the reasoning runs on Claude Haiku in the cloud at a ~20-second cadence, about $1–2/hour, fast enough to stay in the conversation. We tuned the prompts, cadence, and phase logic against real recorded calls with a replay-and-diff eval harness across several industries.

/// a_real_desktop_app
one-click install

Native Windows app

Ships as a single self-extracting .exe: no Python, no setup. A first-run wizard handles the speech model download and the provider key; after that it goes straight to the start screen.

runs beside the call

A native window

A small always-there window renders live transcript, recommendations, and questions: built as a native desktop app, not a browser tab or a web service.

ready before the buyer joins

Preflight + sessions

A pre-call check verifies the key, audio path, and GPU so nothing fails live. Every call is a named session with history and an instant end-of-call summary.

services involved: product development · ai

/// why_it_works

Live help in the call: on your data, on your terms.

01Help arrives in the call, not after it: ~20-second live cadence
02Private by design: the call audio is transcribed on-device
03A reusable template: point it at any data, any sales motion
04Claude or Claude on AWS, and cheap to run at ~$1–2/hour
/// ready_when_you_are

Want a co-pilot like this for your sales team?