How Sharing Your Data Without Thinking Can Cost You Everything

What we talked about.

Sunny Ray talks with Julia, CEO and co-founder of Glender, a company that automatically detects and removes protected health information from medical images, scans, voice, and reports at the source. Julia describes a two decade career spanning mathematics competitions, Berkeley and Cornell, building Wikipedia's did you mean search across 300 plus languages, and pronunciation technology that became part of Rosetta Stone. She explains why AI in healthcare cannot rely on public data the way general chatbots do, since medical records are private and often narrow in scope, leaving most FDA approved AI models unable to generalize beyond the population they were trained on. Julia reflects on being one of few women in early AI, the discipline of spending five plus years building Glender before the market caught up, and the hard call of walking away from startups that were not working. She closes with a warning that once patient data is shared without safeguards, it cannot be taken back, and urges listeners to think carefully about how their data is shared and protected.

Glender's Julia on scrubbing patient data from medical AI before privacy pays the price of progress.

The questions, and the answers.

You grew up around mathematics and competitions. What did numbers give you as a child that people sometimes couldn't?

Math is not only numbers, for me it was geometry, topology and logic, it is so many different things all encompassing. What I loved about competitions is getting together with other people and for three hours solving really difficult problems, so difficult that if you solve three and a half or four out of five you are the winner. It is tough and a lot of fun.

You built the did you mean capability for Wikipedia across 300 plus languages. What was that like?

Wikipedia uses its own bespoke technology and search engine, and they needed help with did you mean, similar to what Google does when someone mistypes or is not sure how to spell something. Because Wikipedia supports 300 plus languages, the algorithms had to be general enough to cover a big swath of them, with different approaches for Asian versus non-Asian languages and plenty of edge cases along the way.

Your pronunciation technology became part of Rosetta Stone. Can you take me to the moment you decided to sell that work?

We built a company called Better Accent to help non-native speakers pronounce things in their second language, converting the prosody of speech, intonation, stress and rhythm, into visible patterns on screen so users get feedback they can see. We had customers in over 15 countries. At some point it stops being about challenging technology and becomes about working with customers, and there are people much better equipped than me for that.

You've said you love work that has no straightforward answer. What scares you about that, and why do you run toward it anyway?

I don't think I'm scared, that's what makes it challenging and exciting. If something has already been done, it becomes a question of optimization. But if you're solving something that hasn't been solved before, nobody has the answers, and that's the exciting part. Some people get their adrenaline skiing off very high hills, which I don't do, but I really enjoy doing this.

Five plus years building Glender before the market was ready. What kept you going through those quiet years when no one was clapping?

I don't think anybody ever claps anyway, so if you're waiting for that, good luck. It's an interesting problem and it needed to be solved. It wasn't just me, it was my entire team, we needed to solve it, so we kept working on it.

Courage is one of your favorite words. What is the bravest decision you've had to make as a founder?

Stopping doing what we were doing. There were a couple of startups that didn't make it, and it's hard to walk away once you realize it's not leading anywhere.

In plain terms, what does Glender do and whose life is quietly different because of it?

Hopefully it helps healthcare in general by solving a problem that is not obvious but extremely painful, connecting data with researchers without impacting patients. With an X-ray, for example, the patient's name and other information is often embedded in the image itself, and that information isn't needed but it blocks the X-ray from being used for research, AI training and other projects.

There are over 1,000 FDA approved AI models, most trained on narrow data. What's the danger you're racing to prevent?

The average number of installations for these models is one, meaning a model trained on a certain population is only applied to that population and nowhere else. If it's trained on data from New York City, it won't be applicable to a tribal clinic in Utah, so these models aren't really transferable across different populations.