Hi, I'm Edoardo: from audio software to a tiny call assistant
Hello Product Hunt. I'm Edoardo, from Italy.
For years I was a product manager on audio-editing and entertainment apps. Later I worked on AI voice products built on real-time speech, including simulated calls used to train people. Two lessons stayed with me: in audio, a one-second delay feels broken; in voice AI, people trust a tool only when they can see what it heard and stay in control.
This autumn I turned those two lessons into my own thing, a small assistant for client calls, and launched it here last week. It was my first launch and I did it alone, with nobody to notify. The ranking mattered less than I expected; the conversations mattered more. People asked whether looking things up during a call is cheating, and whether tools like this make us lazier. Those questions changed how I talk about what I build.
Things I'm happy to talk about: on-device speech recognition, latency, pricing a solo product in Europe, and launching without a network.
If you also moved from a big company to building alone, what was the first habit you had to unlearn?
Replies
Welcome, Edoardo! I only worked for a company for two years, while building my first product on the side, so I didn't have much to unlearn from that. The habit I did have to drop came much later: doing everything myself. These days I make the decisions and hand almost everything else to AI agents. Seeing what it heard and staying in control is exactly how I want my own tools to work too.
Welcome Edoardo! "People trust a tool only when they can see what it heard" applies to music too, in a way I did not expect. In my song app the music model returns the lyrics it planned to sing, and on Russian test songs every one drifted from that plan: a line dropped, a line sung twice, two chorus lines merged. Showing the plan would have been showing what the tool intended, not what it did. So a step now listens to the finished track and the screen shows what was actually sung. Curious about your side: with on device recognition, how do you show the user a word the model was unsure of, without cluttering a live call?