Levers (YC S22)

Levers is a Y Combinator (S22) company building AI for B2B receivables and collections. Finance teams lose huge amounts of time chasing invoices across email, phone, and chat, and slow payments tie up the cash a business needs to grow. Levers automates that whole workflow, including human-sounding voice agents that call customers, answer questions, and help resolve payments. Learn more at uselevers.com.

What I work on

I work across the deep-tech core of the product: the speech systems that let a voice agent understand a caller in real time. Lately that has meant building self-hosted speech recognition, training and adapting models for spoken Arabic and its dialects, exporting and serving them on GPUs, and tuning them for low-latency streaming so the agent can listen and respond as the conversation happens.

Why it is hard

Real phone calls are noisy and fast, full of dialect, code-switching, and domain terms. Off-the-shelf models miss too much, and hosted APIs get expensive and hard to control at scale. Getting accuracy, latency, and cost to line up all at once, under real traffic, is the actual challenge, and it is most of the work.

What it adds up to

Beyond the models, I have led a major architecture revamp and invested in reliability, compliance, and cost-efficiency so the system holds up in production. The unglamorous engineering is what makes voice AI dependable enough to trust with real money.