LightfieldAI-native CRM that builds itself and does work for you
Promoted
Launching today
Ojin
Talk to an AI Agent with a real face and voice, in real time
53 followers
Talk to an AI Agent with a real face and voice, in real time
53 followers
Human AI Agents have a real face and a real voice, and run live conversation rather than turn-based exchange. Interrupt mid-sentence, trail off, talk over it, and the endpointing holds. Setup is one still photo, a persona and a voice. No rig, no capture session, no script. Two face models behind a single API. Portrait for scale, Presence for expressiveness. Both drop into Pipecat and LiveKit. Try to interrupt it. Most demos cannot survive that, and it is the fastest way to judge this one.
Ojin
Hi Product Hunt. I am Mio, founder of Ojin.
I have sat through a lot of AI demos where the face looks perfect but the conversation is unstable.
You say something, it waits. You pause to think, it talks over you. You interrupt, and it finishes its sentence anyway.
Everyone nods and nobody says the obvious thing, which is that this is not a conversation.
So we built Human AI Agents.
What it is
One still photo, a persona written in plain language, and a voice. You get an agent you can talk to and interrupt.
What runs underneath
Two face models behind a single API. Portrait for speed and scale, Presence for expressiveness. Both stream over WebSocket and drop into Pipecat or LiveKit.
What we actually spent the time on
Turn-taking. Knowing when someone has finished a sentence rather than paused to think. It is the unglamorous part and it is most of the product.
Try to break it. Interrupt it mid-sentence, talk over it, trail off, use an accent, most of all have fun!
I'll be around all day.
@iammio Nice launch congrats 🙌what was the single hardest edge case to iron out when tuning the silence detection window?
Ojin
@vikramp7470 Thanks Vikram! The hardest part wasn't tuning any single detector - it was finding one silence-detection approach that held up across every scenario we throw at it. We tested on-device VAD, cloud-based VAD, and external providers like ai-coustics and Deepgram, and each shined in some conditions and fell apart in others. No single method was robust everywhere. What finally worked was a combination of them rather than picking a winner - leaning on different signals depending on the context. Getting that blend to behave consistently was the real grind.