izanton
Joined 2 years ago
Joined 2 years ago
I wrote this article about my attempt to use an external classifier to separate backchanneling and floor-taking intents, but it seems to have too much latency, so it becomes unusable both ways: using built-in VAD makes the experience choppy because any input interrupts the speech, and a custom one takes too much time before it processes the input and interrupt.
So how do you guys solve this problem in production? (Don't suggest waiting for the `gpt-live-x` family. Waiting is killing.
Staff/Senior product engineer with 10+ years of experience building complex web and desktop products. TypeScript is my strongest language. I work across architecture and implementation, especially where rich client experiences, backend systems, and AI workflows meet.
Recent work includes production Electron apps for macOS and Windows; low-latency screen recording with FFmpeg and Web Workers; and a screen-activity recognition system that uses local Tesseract.js OCR as a pre-check, with vision and LLM models for verification. At TheGuarantors, I led the UI kit and helped replace one-person ownership with a frontend committee and a practical contribution path for product teams. I have also reduced a build-and-test pipeline from more than 20 minutes to about 3 minutes.
I am looking for full-time Staff/Senior full-stack, AI product, desktop, or technical leadership roles. I am most useful on products with hard technical constraints and meaningful product ownership. Open to contract-to-hire or longer contract work.
Staff/Senior product engineer with 10+ years of experience building complex web and desktop products. TypeScript is my strongest language. I work across architecture and implementation, especially where rich client experiences, backend systems, and AI workflows meet.
Recent work includes production Electron apps for macOS and Windows; low-latency screen recording with FFmpeg and Web Workers; and a screen-activity recognition system that uses local Tesseract.js OCR as a pre-check, with vision and LLM models for verification. At TheGuarantors, I led the UI kit and helped replace one-person ownership with a frontend committee and a practical contribution path for product teams. I have also reduced a build-and-test pipeline from more than 20 minutes to about 3 minutes.
I am looking for full-time Staff/Senior full-stack, AI product, desktop, or technical leadership roles. I am most useful on products with hard technical constraints and meaningful product ownership. Open to contract-to-hire or longer contract work.
But as I understand it, the most "basic" approach now with qubit-based computers is to select the most popular answer across many runs and treat it as the "right" one.
I know it's sounds stupid, but what if
Socials: - github.com/wapgear - t.me/izmailovdev
---