Insurance Sales Lab
Real sales calls recorded from any meeting tool, never cutting the agent off
Insurance Sales Lab wanted its agents' real calls scored, not only their practice calls. We built a signed Windows and macOS tray app with a virtual microphone that records both sides of any call, with audio routing that fails open so the agent is never cut off.
EdTechDesktop call recorderAI added to a live product
- Status
- In use by the client's agents
- Platforms
- Windows with ARM64, Intel and Apple silicon Macs
- Works with
- Any meeting or calling tool on the laptop
- Product lines
- Insurance Sales Lab and Insurance Growth Lab
The challenge
Real sales calls happen in many different meeting tools on many different laptops, and the client wanted them scored like practice calls without ever risking the agent's live audio.
What we did
We built a signed desktop app with a virtual microphone, Krisp noise cancellation and audio routing that stays on independently of recording and fails open, then hardened it on a dedicated audio test rig.
The outcome
The recorder is in use by the client's agents, and desktop builds for two product lines, Insurance Sales Lab and Insurance Growth Lab, now ship on Windows and Mac.
The starting point
Insurance Sales Lab trains insurance sales agents on its own learning platform, where the AI voice roleplay trainer we built already scored their practice calls. The agents sell on phone and video calls, in whatever meeting tools they already use.
What had to change
The company wanted the same scoring on real sales calls. That meant capturing both the agent and the customer on many different laptops, then transcribing and analyzing every call.
What could not break
- The agent's microphone keeps working even when processing fails
- Both sides of the call are captured, on Windows and on macOS
- Consent comes first, and a privacy stop means nothing from that session is uploaded
- No cookies, tokens or signed URLs ever appear in logs or telemetry
- Audio drivers must be signed to install on customer laptops
The calls we made
We put one capture layer on the agent's own machine, so recording works with any meeting tool, and designed the audio path so recording problems can never cost the agent a live call.
- 01
A virtual microphone on the desktop
An Electron desktop app with a virtual microphone and system audio loopback records both sides of a call in any meeting tool the agent already uses.
Ruled outAn AI that joins each call as a participant, the other option we researched.
- 02
Audio routing that never depends on recording
We moved the virtual mic to always on routing, separate from recording. Stopping a recording or a privacy stop never interrupts the voice, and if processing fails, the raw microphone passes straight through, with a live health check and a notice to the agent.
Ruled outAn audio path that starts and stops with each recording, where a processing failure could silence the agent mid call.
- 03
Reproduce every issue before fixing it
A dedicated audio test rig drives the real app with a simulated meeting tool, scripted microphone reopens and sleep like interruptions, so each reported issue was reproduced, fixed and tested again, with 303 automated tests behind the audio path.
Ruled outFixing field reports by guesswork on one laptop.
- 04
Signed drivers on every chip
The custom Windows audio driver goes through EV code signing and Microsoft attestation signing and was rebuilt for ARM64, with macOS builds for Intel and Apple silicon.
Ruled outAsking agents to install unsigned audio drivers on their laptops.
- 05
Two builds, not one binary
When the client launched a new standalone platform with cookie sessions and its own access tags, we shipped a separate desktop build for it and kept both running in parallel.
Ruled outOne binary for both platforms, which the client ruled out.
How the system fits together
What we built
Tray app with automatic recording
Sign in against the client's platform, automatic start and stop when a call begins or ends, capture of agent and participant audio, and an upload queue that recovers recordings made offline.
Noise cancellation in the live path
Krisp noise cancellation on the virtual microphone, gated by the platform's subscription tags and later by access tags on the client's new platform.
Conversation logs
Recorded calls are transcribed by speaker with AssemblyAI, then AI detects the call type and outcome, so each real call lands in the agent's conversation logs next to practice calls. Owners and team leads open each call in the CallJoyna portal, with its summary, transcript, playback and Ask Joyna for questions about the call.
Signed audio capture
A custom Windows virtual audio driver signed through EV code signing and Microsoft attestation signing, rebuilt for ARM64, and macOS capture through the BlackHole driver and a CoreAudio tap.
Release engineering and test rig
Signed installers that bundle the driver, auto update with release notes, CI/CD from staging to production behind an approval gate, and an audio test rig with 303 automated tests for routing, mic reopen races and power events.
How it works
Detect
The app sees a call start in whatever tool the agent uses and starts recording both sides.
Protect
Krisp cleans the agent's voice on the virtual microphone, and the real microphone stays live if processing fails.
Upload
When the call ends, the recording is queued and uploaded through a signed URL, unless the agent stopped it for privacy or it is too short.
Analyze
The call is transcribed by speaker, and AI detects its type and outcome.
Review
The call appears in the agent's conversation logs, next to their practice calls.
A closer look
{
"call_id": "call-20418",
"captured_by": "desktop recorder",
"speakers": ["Agent", "Customer"],
"call_type": "Cross sell",
"outcome": "Follow up scheduled",
"transcript_status": "complete"
}What changed
The desktop recorder is in use by the client's agents, so real calls are captured and analyzed alongside practice calls. Builds for Insurance Sales Lab and Insurance Growth Lab ship on Windows and Mac, and a separate build keeps agents recording while the client moves users to its new platform.
Under the hood
- Datics role
- Product built by Datics
- Audio capture
- A custom Windows virtual audio driver with WASAPI loopback, macOS capture through the BlackHole driver and a CoreAudio tap, and audio rendered straight into the virtual mic for a steady delay.
- AI services
- AssemblyAI transcribes each call by speaker, and AI analysis detects the call type and outcome for the conversation logs.
- Secrets and uploads
- No cookies, tokens or signed URLs appear in logs or telemetry, and recordings upload to cloud storage through signed upload URLs.
- Distribution
- Signed installers bundling the driver, auto update from release feeds on AWS S3 and Google Cloud Storage, and Sentry error monitoring.
Also in the stack
How it shipped
Research and first desktop app
Research into an AI that joins calls and into a virtual microphone, then an Electron tray app with sign in, meeting recording and Krisp in the audio path.
Signed capture on both systems
The Windows driver through EV and Microsoft attestation signing, native Windows loopback and macOS capture, and the rebrand to CallJoyna.
Reliability on real laptops
Participant audio, echo and distortion fixes, auto update, CI/CD, an ARM64 Windows driver, and Intel and Apple silicon Mac builds.
Client roadmap and a new platform
Always on routing that fails open, a recording status with a privacy stop, a cap on long calls, sleep and crackle fixes, and a build for the client's new standalone platform.
Built to the client's roadmap and rules
The client's own roadmap set the priorities, and its integration rules, such as no cookies, tokens or signed URLs in logs, shaped the build.
Staged releases with a written history
Builds move from staging to production through CI/CD with an approval gate, and each client build came with a plain account of what changed and what was fixed.
Hardening on real laptops
We worked through sleep, USB mic hot plug, echo and crackle on agents' machines, with error monitoring and audio diagnostics in the logs.
What this unlocked
The client's roadmap adds real time two way transcription with live AI feedback to the agent during calls, phone platform integrations, and moving every user to its new standalone platform.
Related work
Have a product with the same kind of constraints?
Tell us what already exists and what you want it to do. We will tell you what it takes.