Empy.ai
December 2021 to May 2026 · closed
It started as a pet project in the month I left Google, December 2021. The idea was narrow and, I thought, testable: most of the damage in a conversation happens when someone states a conclusion as if it were an observation. Nonviolent communication has a name for that distinction. I wanted a model that could hear it.
Two of us spent half a year fine-tuning a transformer to separate evaluation from observation in written text and got it to 91% F1. Two weeks after we shipped it, ChatGPT launched.
Three products in four years
The first was a Slack bot, in 2023. It read team channels, flagged the sentences that carried judgement instead of fact, and offered a rephrasing. Four thousand people installed it across forty-three companies, and it went to Product Hunt three times; the best launch came second product of the day with 203 upvotes. It was also, in hindsight, a product people found interesting rather than one they paid for.
The second was built for remote workers with ADHD: an agenda, focus reminders and engagement tracking inside Zoom, Google Meet and Teams. That is the version that raised the money, a $155,000 pre-seed from two angels, and the version I took to Draper University in the summer of 2024 and to Burning Heroes in San Francisco, where it was one of nine finalists from more than two thousand entries and pitched live at Plug and Play Silicon Valley.
The third was a macOS application that ran during live calls. It showed talk-time balance, engagement on both sides, the unmet need it heard and the question being asked, and it nudged you while there was still time to change course. Seven US enterprise pilots came through it, the largest into an organisation of more than six thousand people; Erickson Coaching International and Facility Solutions Group were among them. By late 2025 it had three thousand people on the list and twenty-three paying.
What it cost
Twelve attempts to ship a product. Eleven co-founders. About two hundred investor calls over Zoom. $40,000 of my own money and $155,000 of other people's. Under $15,000 of revenue across the years it ran. Three accelerators that gave me nothing, and Draper, which did.
How it ended
The Electron client crashed enough to cost retention. In February 2026 I locked myself in a room for three weeks and rewrote it in Swift. That fixed the crashes. It did not fix the business: there was no way to build a real-time harness reliable enough for calls as complex as the ones I had sold to customers. I closed the company in May 2026. The product site is still up at empy.ai.
The diagnosis I had written down five months earlier is the most useful thing the company left me. I had been building a mission instead of a product, and a mission lives in the future, where it can never be shown to be wrong.