Levers (YC S22)

Live at www.uselevers.com

Levers is a Y Combinator company (S22) building voice AI for B2B collections. I help build it as an AI Research Scientist on the team, and my corner is the deep-tech core.

What that means in practice

Collections is a conversation problem before it is a payments problem. Someone has to call, understand what is actually going on, and respond like a person. Doing that with a voice agent means the hard parts are not the business logic. They are latency, recognition accuracy, and the thousand ways a real phone call is messier than a demo.

What I work on

Self-hosted speech recognition, and real-time voice agents built on top of it. Running recognition ourselves rather than renting it is the difference between controlling the latency budget and hoping someone else's API is having a good day. In a live conversation, delay is not a metric. It is the product.

What I have built so far

• 11 Arabic language models, from 4B to 80B parameters, trained and shipped end to end for the Najdi and Hijazi dialects.
• Arabic speech recognition went from a 40% word error rate to under 10%, on hundreds of hours of curated call audio. Next stop: under 5%.
• Helped design and build the whole Arabic voice agent stack from scratch: speech-to-text, LLM, text-to-speech and our own turn detection. No vendor lock-in.
• A direct line into Saudi telecom carriers, so our agents call subscribers on a live national network without a third-party CPaaS in the middle.
• The Levers Dataset Factory, which takes raw call audio all the way to a deployed model: annotation tooling, synthetic data, corpus governance and automatic rollback.
• Turn latency down 85%, from 4 to 7 seconds to about 1 second. On a phone call, that is the difference between a conversation and an awkward silence.

It is the most technically demanding work I do, and it is deliberately separate from Terang AI. One is a company I built to widen access. The other is where I go to get much better at the engineering that makes it possible.