Voio.

It's a Tuesday, and the worklist won't stop growing.

Somewhere in a darkened reading room, a radiologist is on their fortieth study of the shift. A chest CT here, a brain scan there, a breast MRI queued behind both. Each one demands full attention. Each one is a person waiting for an answer. The pile does not care that it is late, and it does not shrink. This is the quiet crisis of modern medicine: not too few images, but too few hours in which to read them carefully.

Voio was built for that room. The Berkeley-and-UCSF spinout, which stepped out of stealth in November 2025, does not promise to empty the worklist by removing the radiologist. It promises something less flashy and more useful - to make the next hour of reading faster, cleaner, and less full of the grunt work nobody went to medical school to do. The company's founder puts the mission in a single question: "How do we help to make it better to be a radiologist next year than it is to be a radiologist this year?"

Three researchers, one stubborn belief.

Voio did not appear out of nowhere. It grew out of the same soil that produced much of modern computer vision. Co-founder Trevor Darrell founded Berkeley AI Research and led the team behind Caffe, one of the deep-learning frameworks that made image recognition practical in the first place. Co-founder Maggie Chung is a practicing UCSF radiologist who knows the reading room from the inside. And chief executive Adam Yala spent years building models like Mirai, a breast-cancer risk predictor validated across more than two million mammograms, and Sybil, which estimates lung-cancer risk from a single screening scan.

What they share is a refusal to accept the usual framing. For years the story about AI in radiology was a threat narrative - the machine comes for the job. Voio's founders read the same data and drew the opposite conclusion. Demand for imaging keeps climbing. The bottleneck is human attention, not human competence. So the useful question is not "how do we replace the radiologist" but "how do we give the radiologist their day back."

Pillar-0, and a workspace that stops making you click.

The lab's first release is Pillar-0, a foundation model that reads CT and MRI scans directly and recognizes hundreds of conditions across the chest, abdomen, brain and breast. In the company's reported benchmark it reached .87 AUC across more than 350 distinct findings - a gap the team measures at 10 to 17 percent over leading proprietary models. The unusual part is not only the accuracy. It is that Voio open-sourced the thing.

Open-sourcing a flagship model is a strange move for a venture-backed company - until you remember the customer. Medicine does not adopt what it cannot inspect. By putting Pillar-0 in the open, Voio is betting that transparency, independent validation and global collaboration will win faster than a locked API ever could, especially in health systems that could never afford a proprietary black box.

How Pillar-0 stacks up.

A tenth of a point on an AUC scale sounds academic. In a reading room it is the difference between a finding caught and a finding missed - repeated across the thousands of studies a department reads in a week.

The people behind the lab.

A short but busy history.

• Before Voio
  Founders build Mirai and Sybil, validated across 90+ hospitals in 30 countries and 2M+ mammograms.
• Nov 2025
  Voio exits stealth with $8.6M seed and open-sources Pillar-0, calling it the most accurate public radiology model to date.
• Apr 2026
  UC Berkeley and Berkeley News feature the team's work using AI to revolutionize medical imaging.

See it for yourself.

Interviews, launch coverage and the Pillar-0 project page.

The worklist is still long. The Tuesday is different.

Return to that reading room. The pile has not vanished - imaging demand does not work that way, and Voio never claimed it would. But the fortieth study of the shift arrives with a draft already attached: findings surfaced, the routine measurements taken, the grunt work quietly done. The radiologist reads, corrects, decides. The judgment stays human. The clicking does not.

That is the whole of Voio's argument, stated plainly. Not a machine that replaces the reader, but a tool that makes it better to be one - next year than this year, and the year after that. A small team out of Berkeley and UCSF put a very accurate model into the open and asked the field to check their work. Whether the future of radiology looks like their pitch is not yet settled. But the room is a little quieter, and the answer comes a little sooner. For the person waiting on the other end of the scan, that is not a small thing.