Dylan Fox

Dylan Fox runs AssemblyAI, an AI research company in San Francisco whose Speech AI models sit quietly behind products people use every day. When a note-taking app transcribes your meeting, or a call-center tool tags the moment a customer got frustrated, there is a decent chance Fox's software did the listening.

His pitch has stayed unusually consistent since 2017: take the kind of speech models that once lived inside big research labs, and hand them to any developer through a single API. AssemblyAI now processes more than 25 million inference calls a day and counts over 200,000 developers on the platform, with paying customers ranging from Spotify and Zoom to Fireflies.ai and CallRail. The company has raised roughly $115 million, most of it in a compressed 22-month stretch that ended with a $50 million Series C led by Accel in December 2023.

What makes the arc worth reading is how slow it was at the start. Fox spent about three years grinding to his first million dollars in revenue before anything looked like a rocket ship. The lesson he tends to draw from that is not glamorous: endurance, not luck, is what kept the company alive long enough to catch the wave.

A complaint that became a company

Before AssemblyAI, Fox was a machine learning engineer at Cisco, working on collaboration products that kept bumping into the same wall. Whenever a project needed speech recognition, the options on the market felt stuck in an earlier decade.

All of the companies offering speech recognition as a service were insanely antiquated, hard to buy anything from, and were running outdated AI tech.

Dylan Fox, on the problem that started AssemblyAI

He had come to machine learning sideways. Fox studied at George Washington University and learned to program in part by showing up to Python meetups around Washington DC. College coursework pulled him toward algorithm-heavy problems, and those problems pulled him toward machine learning and natural language processing. By the time he was at Cisco, the pattern was set: he liked the hard, mathematical edge of the field, and he could see a gap in the market that nobody serious was filling well.

So he quit. Not with a co-founder and a polished deck, but as a solo founder with a prototype and a conviction.

The late application

The origin story that follows Fox around is the Y Combinator one. He recorded a video explaining his technology and submitted an application to the accelerator roughly 30 days after the deadline had already passed. Instead of a form rejection, he got an interview.

At the YC office he met Daniel Gross, a former Apple employee who had looked hard at speech recognition and immediately understood what Fox was trying to build. Gross became his first investor. That early belief mattered, because the wider market was not convinced.

When Demo Day arrived about three months later, the venture reaction was blunt. Investor after investor passed with a version of the same line: Google will do it. Rather than argue, Fox raised around $1 million from angels, kept the team deliberately small, and got back to building. It was a bet that the incumbents would move slowly and that developers wanted something better than what the giants shipped.

Our mission is to make state-of-the-art AI models accessible to developers and product teams at extremely large scale through a simple API.

Dylan Fox, on the AssemblyAI mission

The Twilio-for-speech idea

The mental model Fox reaches for is Twilio, the company that turned telephony into a few lines of code. He wanted to do the same thing for speech: wrap constantly improving research inside an interface simple enough that a developer could ship a voice feature in an afternoon, without understanding acoustic modeling or hiring a research team.

That framing shaped who AssemblyAI built for. The customer was never the CIO first. It was the developer writing the integration, and the product team deciding whether transcription was accurate enough to trust in front of users. Win those people, the theory went, and the enterprise contracts would follow. They did. Companies including the Wall Street Journal, NBCUniversal, Spotify and Zoom now route audio through the platform.

To keep the underlying models improving, Fox assembled a research team that pulls from alumni of DeepMind, Google Brain and Meta AI. The scale of the training effort has grown accordingly. The company's newer Universal-class models have been trained on more than 10 million hours of audio, which works out to over a petabyte of voice data. Fox frames the goal in deliberately ambitious terms: not merely accurate transcription, but superhuman speech models that can hear and understand better than a person taking notes.

Past the benchmark

Fox has grown skeptical of the industry's favorite yardstick. For years, speech systems were ranked almost entirely by word error rate, the percentage of words a model gets wrong. His more recent writing pushes past that single number, arguing that raw word accuracy misses much of what makes voice data useful: who spoke, what they meant, where the important moments were.

It is a fitting stance for someone who started the company because the existing tools felt too narrow. The through-line across nearly a decade is the same instinct that made him quit Cisco and file a late YC application: look at the standard answer, decide it is not good enough, and build the thing you wish existed.

Through college courses, I found myself leaning more into algorithm-type of programming problems, which naturally led me to machine learning and NLP.

Dylan Fox, on finding his way into the field

None of it looked inevitable in 2017, when the polite consensus was that a bigger company would eventually swallow the whole idea. Nearly a decade later, the developers kept choosing the small company that answered late and shipped early. That is the quiet argument AssemblyAI keeps making on Fox's behalf, one API call at a time.