Kintsugi.

It is a Tuesday afternoon in Berkeley, and a phone is ringing inside a hospital call center somewhere in the central United States. A woman picks up. She has called about a refill. The agent asks how she has been sleeping. She talks for thirty seconds about her week. Somewhere behind that exchange - quietly, invisibly, with the bedside manner of a librarian - a piece of software made by Kintsugi has already heard the answer she did not give.

The model does not record what she said. It listens to the way she said it. Pitch. Pause. Cadence. The shape of a breath caught at the end of a sentence. From those acoustic crumbs it computes a score: the probability that she is sitting on the edge of clinical depression or anxiety. The agent's screen quietly suggests an extra question. A nurse calls back later that day. Two weeks after that, she is in a therapist's office for the first time in her life.

That, in one paragraph, is the entire reason Kintsugi exists.

Voice as a vital sign.

Most of medicine is run on numbers - blood pressure, A1C, white cell count. Mental health is run on a clipboard. The PHQ-9 and GAD-7 are good instruments, but they require a patient who is willing, present, and self-aware enough to fill them out honestly. A lot of people are none of those things on a bad day.

Grace Chang and Rima Seiilova-Olson founded Kintsugi in 2019 with a different proposition. If depression and anxiety produce reliable changes in how people talk, then twenty seconds of free-form speech should be enough to flag them. No questionnaire. No appointment. No self-report. Just sound.

The company is named after the Japanese craft of repairing broken pottery with gold lacquer. The metaphor was almost embarrassingly on-the-nose, but the founders never apologized for it. They had both lived through periods when mental health care was almost impossible to access. They believed the repair could be beautiful. They also believed it could be measured.

An API the size of a screening question.

Kintsugi's flagship product, Kintsugi Voice, was sold the way nobody glamorizes: as an API. Its job was to sit quietly inside someone else's workflow - a telehealth visit, a remote patient monitoring app, a call center handling refill requests - and return a probability score after twenty seconds of audio. There was no chatbot. There was no avatar with a soft voice. There was a JSON response, and a clinician on the other side who had been given a reason to ask one more question.

Around it the company built two sibling products. Kintsugi Connect tried to solve the awkward second act: what happens after a patient screens positive. Kintsugi App was patient-facing, a journaling tool that turned the same listening model on the user themselves.

Two engineers who refused to handwave.

Grace Chang came out of the Bay Area's signal-processing-for-authentication world before turning that toolkit on the human voice. Rima Seiilova-Olson was a former competitive programmer and ML scientist who had productized models in less consequential domains and decided to spend the next ten years on a more consequential one. Both founders described the company's origin in the same direct register: they had each, at different points, tried to get mental health care and found the system unworkable. They thought they could build a sharper instrument.

Eight years of quiet, in chronological order.

The strangest, most generous part.

In February 2026, Kintsugi announced it was shutting down. That, on its own, is a normal piece of startup news. What followed was not. Instead of selling the models to a larger health-tech buyer, instead of letting the IP sit on a shelf and depreciate, Chang announced that Kintsugi would release its voice biomarker models, scientific methodology, and research into the public domain. Every model. Every method.

You can read the move two ways. The cynical version: the market for voice biomarkers in healthcare had not matured fast enough to support a standalone company, and a public release was a graceful way to land the plane. The generous version: the founders had always said the goal was access, not enterprise value, and at the end they actually behaved that way. Both versions can be true. The interesting thing is that the second one is at least defensible. Most founders, in most exits, give you no version of that story at all.

Where to read more.

Back to the call center.

The woman who called for a refill on that Tuesday afternoon does not know any of this. She does not know the company's name. She does not know about the Series A, or the FDA submission, or the office on Solano Avenue, or the eventual decision to give the model away. She knows that two weeks ago she was sitting on the floor of her bathroom, and now she is sitting in a therapist's office, and the line between those two things ran somehow through a phone call she barely remembers.

Kintsugi the company is now in past tense. Kintsugi the technology is sitting in a public repository, available to any health system curious enough to deploy it. The call center is still ringing. The model is still listening. And the question Grace Chang built her company around - whether twenty seconds of ordinary speech can be enough to catch a person before they fall - has finally stopped being a thesis, and started being a piece of infrastructure that nobody owns.

— END —