A private live coach for video calls
A real-time call coach that runs on your laptop
A Mac app that sits around the notch and coaches you live during a video call — surfacing your talking points, flagging when you're talking too much, reading the room — without sending anything off your machine.
People go into video calls with a goal and routinely miss it for the same few reasons — forgetting a prepared point, talking too much, misreading the room — and a live coach that could help wasn't feasible to build.
On-device AI removes the three blockers at once — lag, cost, and privacy — so a glance-able coach can read the call in real time and nudge you, entirely on your own laptop. The work was proving that's genuinely feasible before building it for real.
| Before | After | |
|---|---|---|
| Where the AI runs | in the cloud, on a server | entirely on your own laptop |
| Speed | lag that breaks 'real-time' | fast enough to coach you live |
| Ongoing cost | a per-minute charge for every call | none — it runs on hardware you already own |
| Privacy | a private call streamed to a server | nothing leaves the device |
| Setup | per-app integrations and audio plugins | one app, using the system's own recording path |
Running the models locally turns an infeasible product into a feasible one. The blocker was never the idea — it was lag, cost, and privacy, and on-device AI dissolves all three at once. This is a whole product category that only became possible in the last generation of AI models. The work here was proving that, rigorously, before anyone wrote production code.
What I built
A serious Stage-1 feasibility study — not a finished product, but a rigorous go/no-go on whether a real-time call coach can actually exist. Three things all have to be true for it to work, and the study tested each one honestly:
- Can it run on the device? The technical core is a small Mac app that sits around the notch, wakes up on its own when you enter a call (with no per-app integrations), and captures both your audio and the other person's through the operating system's own screen-recording path — deliberately avoiding the audio plugins that make setup painful and scare people off.
- Do the costs work? Running the AI locally means there's no per-minute charge for every call, which is what makes the economics hold up at all.
- Is getting started effortless? Every permission the app needs was mapped in the exact order a new user would hit it, because each extra permission request roughly halves how many people finish setup.
Once running, the coach surfaces your prepared talking points, flags when you're dominating the conversation, and offers a read on the other side's mood — all live, all on the machine.
Why it matters
The benefit is a private coach that helps you in the moment, on calls that genuinely matter — and the reason it can exist at all is that the AI runs on your own laptop. In the cloud, real-time call coaching is a non-starter: the lag breaks the "real-time" promise, the per-minute cost breaks the economics, and streaming a private conversation to a server breaks trust. Run the same intelligence locally and all three objections disappear at once.
That's the point of the study. It's a clear example of a product the previous generation of AI simply could not support, and this one can — and the valuable work was proving that rigorously, so anyone building it knows it's real before they spend a dollar on production.
The real work here wasn't the coaching — it was honestly answering whether this product could exist at all. So this was treated as a serious go/no-go study, not a build: three things all had to be true at once — it has to run fast enough on the device, the costs have to actually work, and getting from install to first value has to feel effortless. The trickiest of those was the effortless part, because every permission a Mac app asks for on first launch roughly halves the number of people who finish setup. So the capture method was chosen to lean on the operating system's own recording path rather than per-app plumbing, and every permission was mapped in the exact order a new user would hit it.
This is a clean example of a product the previous generation of AI simply could not support, and this one can. The deliverable was the honest feasibility call — proving the idea is real, and showing exactly why on-device AI is what makes it possible.
The design for waking up, capturing both sides of the audio, and requesting permissions without per-app integrations is the hard-won part; the full feasibility study is available to walk through.