Hi, it’s Phonic.

Imagine that you’re building a voice AI app, Sylvie, that handles your restaurant’s phone line. The vision is clear—seamless, lifelike conversations that actually help your customers. But behind the scenes, it’s full of compromises.

One morning, you reviewed some of Sylvie’s recordings and found that on one, she launched into an impassioned, off-topic tirade about the secret recipe of her favorite quiche. Moments later, another recording revealed her confidently directing a customer to a made-up address.

Guiding a conversation towards some task-driven goal is essential for the reliability of voice agents, but the entire industry is currently struggling with this. To solve your problem, you try and constrain Sylvie into moving between pre-defined states, but she couldn’t handle the fluid reality of conversations: follow-up questions didn’t fit in any particular state, and you felt like you are playing Whack-a-Mole by consistently adding in edge cases.

What Sylvie needs isn’t more constraints — it’s understanding, true conversational intelligence that guides without shackles. At Phonic, we believe in the next evolution of Voice AI, one that doesn’t hamstring the AI’s intelligence.

Enter Phonic: Where voice AI finally grows up

Phonic is a speech-to-speech platform designed to reliably handle task-oriented workflows. No tying APIs together, no clunky state machines. Just feed your audio in, get audio out, with 300ms end-to-end latency, conversational voices and an agent that does what it should.

How? We threw out the rulebook. To address realistic and fast voices, we trained and post-trained our own models with a focus on reliability. To address reliability, we surrounded them with compound AI systems that manage and guide conversational state seamlessly. Hear some samples for yourself:

Phonic learns and improves from every call, pulls in previous calls and documents, and asynchronously guides the conversation with more intelligent but slower models to improve reliability while not hurting latency:

The future is a workflow, not just a model

We’re not here to sell you another model. We’re the entire platform for Voice AI: we’ll help you build your voice agents, observe them, and evaluate them. For our customers, we’re their system of record.

One of our design partners is building an operational layer for voice-driven healthcare assistants. Before working with Phonic, they struggled with maintaining their state machine across thousands of calls and navigating complex code. After Phonic, they were able to delete significant complexity in their codebase, experience higher voice quality, and significantly better reliability.

Why us?

Our team has worked at institutions like MIT, Stanford, MosaicML, Meta, and Genesis Theraputics, and contributed to machine learning research in both academia and industry.

We’re based out of San Francisco and believe that our best work is done in-person. We enjoy our craft, but also have a lot of fun along the way. If this sounds like something that resonates, then please reach out! We’d love to talk to you.

We’re backed by $4M from Lux Capital and visionary leaders in the AI space including the founders of Replit (Amjad Masad), Hugging Face (Clem Delangue), Applied Intuition (Qasar Younis), and Modal (Erik Bernhardsson). And we’re just getting started.

Your move

If you’re tired of patching leaks in your voice AI raft, let’s build a speedboat. The era of “good enough” voice AI is over. Let’s set the new standard together.