PulmoForge

• 

  Homepage

• 

  Technical Articles

Problem and Inspiration

AI deepfake fraud is causing billions of dollars of losses globally and projected to cause $40 billion in losses by 2027. The biggest percentage of it is voice deepfake fraud, surging 1,300%, causing an average of $500k in losses per incident and continuing to grow. The problem with other detection companies is that they are either expensive or enterprise systems, making most organizations or people vulnerable to just accepting the risk. The reason for the inspiration is that I saw all of these risks and an open market, and when I got into this project, I found this sector actually interesting.

Solution

PulmoForge is a real-time, lightweight AI voice detection system designed to be across communication channels and eventually be accessible on individual phone calls. It is currently trained on large amounts of diverse real and fake audio with differential conditions and built to be a reliable alternative to expensive or enterprise systems.

What Did We Build?

I built and trained the current version of PulmoForge as an SSL and self-supervised model that is trained on large-scale commercially viable datasets in different conditions, like meeting compressed audio, room reverberation, and more.

The Main Challenge We Faced

The main challenge my cofounder and I faced is getting viable customers. This is definitely a validated market, with billions of dollars of losses, but we didn't know how to handle customer outreach. We tried in-person cold pitching to financial institutions, calling law firms and insurance companies, and emailing other companies, but the only responses we got were laughter, hang-ups, refusals, and no responses. Right now that is still one of our main challenges.

What we are proud of

At the start of this, I just had a vibe-coded prototype and never thought I could master this. I thought the market was going to be hard, but I mean, I kind of understand it partially now.

Things we learned:

Audio security, cold pitching, machine learning, partial Python, TTS systems, basic neural networks, training models, and more.

What's next for PulmoForge

Continuing to expand and work on the project

Built With

• audio-security
• google-colab
• kaggle