Meeting AI is solving the wrong half. Here's our fix.

A few years ago I lost a deal in the final five minutes of a meeting.

I knew the answer. I had prepped for it. There was a whole framework in my notes for exactly the objection that came up. But under the pressure of the room, with the camera on me, with the prospect waiting, I blanked. I gave a soft, vague reply. The room felt it. We never recovered.

I don't think I'm unique. Everyone whose work happens in conversations has lost a moment that mattered because the right thing didn't make it out of their head in time. The maddening part is that the right thing was there: in their prep, in their muscle memory, in their notes from a year ago. It just didn't surface.

That gap is what we're building IntelliPrompter to close.

There are dozens of meeting AI tools today, and I've used most of them. They share a model: record the meeting, transcribe it, send you a summary an hour after it ends. Fathom, Fireflies, Otter, Granola. All of them valuable. All of them solving the wrong half of the problem. They tell you what happened. They can't change what happens.

That bothered me enough to build something different.

IntelliPrompter is live meeting intelligence that runs during the meeting. Quietly, on your screen only. It listens to the conversation as it unfolds and surfaces the script you prepared, the framework that applies, the question you should ask, the fact that disproves their objection. While the room is still yours to shape.

Real-time, not retrospective. Post-meeting summaries are a commodity. Anything that doesn't help you in the moment isn't pulling its weight. Every feature has to earn its place in a sub-second response budget.

Privacy by architecture, not by promise. Your highest-stakes conversations are the last thing you want uploaded to a vendor's servers. Transcription runs locally on your machine via Whisper. AI calls go through your own provider keys, or fully offline through Ollama. The trade-off between cost, capability, and what leaves your device is your call. We built the product so we couldn't be the company that leaked your renewal negotiation, because we never built the path.

Model-agnostic. The best model today won't be the best model in six months. Smart routing picks the cheapest capable model for each task. Small classification calls run on local Llama for free. The hard reasoning passes go to Claude or GPT. You get the AI you need, not the AI we signed a deal with.

We're in private beta. It's not perfect. We ship a lot, break things, fix them fast. If you want to be part of the first cohort, the waitlist is at intelliprompter.com/download.

I'll be writing here as we go. Future posts will dig into the specific features. Pain and gain, with the engineering decisions visible. Why a voice-tracked teleprompter actually requires fighting the audio pipeline. The timescale bug that silently broke our coaching engine for weeks. What it cost (literally) to hit our 5–10× AI savings target. The privacy choices we agonized over.

If meetings are where your work lives, I think you'll like what we're building. Welcome.

Jeff