Launching today

uRun
Real-time decisions for voice and video agents
115 followers
Real-time decisions for voice and video agents
115 followers
Sense is a real-time decision model for voice and video agents. It runs on streaming audio and video and answers your questions about the conversation many times per second: is the user done speaking, did they just interrupt, should the agent stop talking. You write the questions, Sense returns the decisions, and your agent acts on them. Build with the SDK for real-time sessions, or send single requests to explore. New accounts start with 1,000 in free credits.






Free Options
Launch Team

Cal.com v7Cal.com now has events, discover, an assistant + more
Promoted
Hey Product Hunt, I’m Keegan, founder of uRun.
Today uRun opens to everyone, and the first model on it is Sense.
If you’ve built a voice agent, you know the timing problem. Your user pauses to think and it jumps in. Your user says “mm-hmm” and it stops talking. Silence timers and rules never quite fix it.
Sense is the first decision model trained to walk forward through a live stream, and the first fast enough to orchestrate turn-taking in real time. It hears the conversation directly and answers your questions up to 12.5 times a second, so your code decides when to speak.
Sign up for free credits and grab the Python or TypeScript SDK: github.com/urun-sh/sense-sdk
Launch post: urun.sh/blog/introducing-sense
Which turn-taking cases break your agents? The whole team is here all day.
Does it work well with different accents and languages too? Love the animated logo!
@bengeekly Thanks! We tried it with Italian and French but can't vouch for our accents - please give it a try and let us know!: https://pypi.org/project/urun-sense
@bengeekly It does! This is just one turn taking benchmark, and some languages are much weaker than others, but the model is fully multilingual and SOTA at turn taking in a number of languages especially when tuning the threshold to match the unique cadence of the language!
The idea of a decision model for continuous perception is mind-bending. How does it work at a high-level?
@shcallaway At a high-level we trained a multi-modal model optimised to answer questions in real-time - so its NOT speech to text, then put the text through an LLM, then wait, then decide - it works straight off your voice (or text, or video).
@shcallaway Double-clicking on what @mpupilli said, at a high level we have trained + optimized a decision model with a couple unique characteristics:
1. The model is able to make decisions in under 80ms
2. The model is trained to make decisions continuously over streams of audio, video, and text/structured data rather than a single shot
3. The model is aware of "now", and what is happening in the current 80ms window of inputs, while still being aware of what has happened before
The combination of these allows you to define a set of questions and criteria for answering, and subscribe to a stream of answers with calibrated probabilities, that you can set thresholds on to trigger events and take actions when things change. This allows you to perform a wide range of useful tasks that used to require specialized smaller models to be able to meet the latency requirements. Things like turn-taking, sentiment analysis, moderation, and computer vision tasks over live streams of video, voice control, gesture-based controls, the sky is really the limit. We're really excited to see what people build with Sense!
The timing problem is so real — every voice agent I've tried either jumps in while I'm still thinking or waits so long the conversation feels broken. A model that decides continuously over the stream, in under 80ms, instead of transcribing first and then asking an LLM, feels like the right layer to fix it at. I'm curious which turn-taking case breaks most often in practice — interruptions, or the long thoughtful pause?
@jcbuilds In my experience its the long thoughtful pause. Its easier to tune an agent to jump in on every pause than it is to handle natural pauses well without losing responsiveness the rest of the time
Amazing work @uRun team , and congrats @keeganmccallum3 ! Upvoted
@keeganmccallum3 Thank you @alejandroesquivel !
@alejandroesquivel Thank you!!
Awesome and congrats on the launch. I tend to just retell for stuff like this and it generally works fine but you definitely can tell it's not quite there. Super excited to try this out
congrats!!!
Thank you @marty_kausas2 !! Appreciate your support