Connie

Inspiration

We kept noticing the same pattern in our own lives: uploading a resume here, pasting financial info there, trying out yet another AI subscription without really tracking what it costs or where the data ends up. Nobody sits down and thinks it through, it just happens in the moment because the tool is right there.

We wanted something that catches that moment instead of judging it after the fact. Not another dashboard nobody checks, but something that actually feels like it's looking out for you. That's how Connie started. The first version of the idea was called "AI Nutrition Facts," basically a label for what's in an AI interaction, and it grew from there into a companion that actually lives in your browser with you.

What it does

Right before you upload a file to an AI tool, Connie steps in and tells you, in plain language, what's in it that you might not want to hand over. She also keeps an eye on what you're spending across your AI tools and whether you're even using the right one for what you're trying to do.

She's built around three sliders, so you decide how much she should push:

how careful she is about your data
how strict she is about your spending
how pushy she is about tool choice

How we built it

Three pieces had to work together for this to feel fast and not annoying:

Jev, from TypeSafe AI, does the actual checking. It doesn't write anything, it just answers a narrow question like "does this file contain personal information" and gives back a verdict with a confidence score. Roughly $P(\text{sensitive} \mid \text{file}) \geq \theta$, where $\theta$ shifts depending on how careful the user's slider is set. Because it's not generating text, it's cheap and quick enough to run on every single file.

Gemini takes whatever Jev decides and turns it into something a person would actually want to read, Connie's actual voice, instead of a flag or a warning label.

Snowflake keeps the history, so you can ask Connie something like "what have I shared with AI this week" and get a real answer.

We split up the work along those same lines: one of us on Snowflake and the backend, one on Gemini and the advice logic, one on the extension and the front end.

Challenges we ran into

Our first idea was to scan what people typed into AI chat boxes in real time, but that felt invasive in the exact way we were trying to prevent, so we scrapped it and moved to scanning files at the point of upload instead. That ended up being a better call.

Keeping Jev "dumb" on purpose was harder than expected. It's tempting to let the same model do the detecting and the explaining, but that muddies both jobs. Splitting them apart was the right call, it just took some back and forth to actually commit to it.

Getting the three sliders to feel like they worked together instead of contradicting each other took a few tries too.

Accomplishments that we're proud of

Getting Jev, Gemini, and Snowflake actually talking to each other in time, that alone felt like a win given the hours we had.

We're proud that the file-scanning approach ended up feeling useful instead of like surveillance, that was the thing we were most worried about getting wrong.

And honestly, getting Connie's voice right, so she reads as a friend and not a warning label, took more iteration than we expected but was worth it.

What we learned

Splitting detection and explanation into two separate models turned out to be the right instinct, a fast narrow classifier plus a model that's actually good at talking to people.

We also learned that you don't need to scare people to get them to care about privacy. Calm and friendly works better than alarming.

What's next for Connie

More categories for Jev to catch beyond personal info, real spending analytics across everything someone's subscribed to, and smarter recommendations for which tool actually fits the task at hand.