AI 現在能回答任何面試問題，我們因此為銷售通話打造即時助手

In 2021 a dev.to author caught a candidate by the white glow of a search page on their face. In 2026 that tell is gone. Overlay tools listen to the call, send the question to an LLM and show the answer in a window that screen sharing can't see. The candidate just reads it out.

Hiring teams are reacting the way you'd expect: more in-person rounds, harder follow-up questions, detection software. It's an arms race, and I don't think detection wins it.

Where the line actually is

The technology isn't the problem. Hiding it is.

An interview tests what you know. If an AI answers for you and the interviewer doesn't know, you're misrepresenting yourself.

A sales call is different. Nobody expects a rep to remember 100 pages of product docs, pricing rules and integration details. Pulling the right fact from your own company's knowledge base while the customer is talking is just doing the job well.

Same technology, opposite ethics. So we built a real-time assistant for the second case, and we say plainly that it's not for interviews.

How it works

The architecture is simple on paper:

A local agent captures call audio on the user's machine.
Audio streams to Deepgram for real-time transcription.
The transcript and the relevant part of the knowledge base go to a cloud LLM.
A hint appears on the user's screen while the call is still going.

The knowledge base has two layers. The shared one holds company-wide material: product docs, pricing, objection handling. The personal one holds the rep's own notes. In our current setup that's around 100 pages of text per company.

Why native C++ and not Electron

A call assistant runs for hours next to Zoom or Google Meet, which already eat a lot of CPU and memory. Shipping another copy of Chromium on top of that felt wrong. We wrote the agent in C++ so it stays light, starts fast and captures audio close to the OS instead of through a browser layer.

The trade-off is real: more platform-specific code and slower UI iteration. For a tool that runs all day in the background, resource use mattered more to us than UI speed.

What we're still figuring out

Latency. The hint has to arrive before the moment in the conversation passes. Every step in the chain adds delay.
Relevance. A wrong hint is worse than no hint. Retrieval quality matters more than model size.
Disclosure. Should the other side of the call know the rep has an assistant? We have opinions, but I'd like to hear yours.
The point

If AI can answer anything, the useful question isn't how to catch people using it. It's where using it is honest.

If you're working on something similar, or think we drew the line in the wrong place, I'd like to hear about it in the comments.

Disclosure: I'm building realtimer.ai