Yoav Zimmerman built a neural network that writes music, shipped machine learning at Google, and grew a startup past a million dollars in a year. His newest company, Third Chair, points all of it at a quieter problem: the rights people own but never have time to enforce.
Somewhere on the internet, an ad is running with your song underneath it. You never licensed it. You may never find out. And even if you did, chasing it down would cost more in lawyer hours than the ad is worth. So it plays. This is the small, constant leak that Yoav Zimmerman decided was worth a company.
Zimmerman is the CEO and co-founder of Third Chair, a company that joined Y Combinator's Spring 2025 batch. The pitch is plain enough to fit on a business card: help the people who own music, film, and other content find where it is being used without permission, then turn those uses into money instead of losses. An AI agent does the searching and the first drafts of the paperwork. A human lawyer stays in the loop.
It is a legal product built by an engineer who spent years far from law. To understand why he is the one building it, you have to go back to the music.
Before the startups, Zimmerman was a machine learning engineer. He worked at Google and later at Determined AI, an open-source platform for training models at scale. His public code from that era reads like a resume for someone who genuinely enjoys the work. There is a transfer-learning experiment. There is a browser-based drum machine he called beatbox. And there is music_rnn, a project that used recurrent neural networks to compose music, one note at a time, which strangers on GitHub have starred nearly two hundred times.
Keep that project in mind, because the name he eventually chose for his legal company is Third Chair - a reference to where a musician sits in an orchestra section. The person who once trained a network to write melodies now runs a business whose biggest early customers are in music. The interests never really separated. They just kept circling each other until they met.
In 2020, Zimmerman co-founded Trendpop and took it into Y Combinator's Winter 2021 batch. Trendpop pointed machine learning at short-form video, reading the patterns behind why some clips take off and others disappear. It was built for the people whose job is to make things go viral: labels, agencies, and creators.
The customer list was unusually broad for a young company. Trendpop worked with Universal Music Group and with MrBeast, with Atlantic Records, United Talent Agency, and Pearpop. The numbers moved fast. Trendpop went from no revenue to more than a million dollars in annual recurring revenue, and to profitability, in under a year. It was later acquired by Collab.
The run earned him a spot on the Forbes 30 Under 30 list for Marketing and Advertising in 2023. It also taught him a lesson he would carry into the next company: the fastest way to a real business is to find work people are already losing money on, and do it for them.
Zimmerman's two companies look different on the surface. One predicted virality; the other enforces copyright. Underneath, they run the same move. Both start with a task that is important, tedious, and quietly expensive - the kind of task that gets skipped because no one has the hours. Then they aim machine learning at it and turn the skipped work into an outcome the customer can see.
At Third Chair, that outcome is money recovered. The company describes its product as a loop with three parts, and it is worth walking through because it explains the whole business.
A rightsholder points the agent at their catalog. It scans social platforms for uses of that content. When it finds an unauthorized commercial use, it does the first pass of the enforcement work - the outreach, the draft demand letter - with a lawyer reviewing before anything goes out. What was a cost center becomes a channel.
Sales has software. Marketing has software. Support has software. Legal, for most rightsholders, still runs on people reading spreadsheets and deciding which fights are worth the billable hours. That is the gap Third Chair is built around. Its own tagline puts it directly: "Unblock Legal, Unlock Growth."
The customers are the ones with the most rights and the least time to police them - record labels, music publishers, distributors, investment funds that own catalogs, and the artists and managers underneath them. One buyer at a top music company, quoted anonymously by Third Chair, put the contrast with older tools plainly: most vendors just hand over data files, while this one does something with them.
This is the part that makes the product easy to sell. Zimmerman is not asking a customer to imagine a future. He is telling them that revenue they already earned is sitting on the table, and offering to go collect it. The conversation ends with a check, not a demo.
There is a version of Zimmerman's career where he stops after Trendpop. An exit, a Forbes badge, a soft landing at a bigger company. Instead he did the harder thing, which is to begin again from a blank repository, this time with a new co-founder in Shourya Lala, himself a repeat YC founder who previously built and scaled the investing app Fello.
The willingness to keep starting over is the trait that shows up most in his work. He tends to arrive at ideas a little early - machine learning infrastructure before it was fashionable, short-form video analytics before the platforms owned the culture, AI agents for legal work before the category had a settled name. Being early is uncomfortable. It is also where the unclaimed problems are.
Here is the piece worth borrowing, whatever you build: the most defensible AI product is rarely the one with the cleverest model. It is the one wired into a workflow the customer already loses sleep over. Build toward the outcome, not the demo. Zimmerman has now run that play twice, in two different industries, and it has worked both times.
The through-line is easy to miss if you only read the company descriptions. An analytics tool, then an enforcement agent, do not obviously belong to the same person. But zoom out and the shape is clear: an engineer who loves both machine learning and the messy business of media, working the border between them, refusing to pick just one.
Third Chair is the current expression of that. If it works, a lot of quiet leaks stop being leaks and start being revenue. And the person who once taught a computer to write a melody will have spent his thirties making sure the people who make the music actually get paid for it.