Logan
Eberwein

Founder

My goal with WellLens is the perfect fitness app, with none of the problems the others have. You answer a few onboarding questions and it builds your workout plan. Then the gym buddy feature keeps you showing up. Every day you work out, you take a quick photo as proof, and your friend holds you accountable.

I started it with a friend. He left the project early on, and I kept going.

It's in beta on TestFlight with my family right now, and I plan to release it in the next few weeks.

Side project

I used to procrastinate a lot, and Checkback fixed that. A parent assigns a task, and the kid sends a photo as proof: a finished chore or a screenshot of their grades for the day. Approved tasks earn points the kid can trade for rewards the parent picks. I've been more productive than ever since I started using it.

It's a much smaller app than WellLens, and I plan to release it in the next few days.

Independent researcher

In August 2025, as a freshman, I started running long ChatGPT and Claude sessions aimed at the Riemann Hypothesis, one of the most famous unsolved problems in math. The main target was an operator whose eigenvalues line up with the zeros of the zeta function. Unsurprisingly, none of the candidates worked, but the point was to create a map of failed approaches with reasoning behind every failure.

I didn't do the math by hand, and the papers say so. My job was direction. I kept the AI going with ideas like these:

• Hit a wall? Build the mathematical system that breaks it. If that system has its own wall, build another one. Keep going up until the top wall breaks, and every wall below it collapses.
• Try, fail, and figure out why. Then try again with what you learned. The failures compound, and so does the shape of the solution.

After a year, that work adds up to 190 documented approaches, 297 operator candidates across 124 cycles, Lean 4 formalization, and high-precision Python computations.

What I learned about AI and math

AI is good at looking like it made progress. A few entries from my own corrections log:

• Lean accepted several "theorems" shaped like A → True. They compiled, and they proved nothing.
• The first version of my report gave μ(2)·log 2 as an example of a positive number. μ(2) is −1, so it's negative.
• One lemma got the name "GNS positivity." All it showed was that a squared absolute value can't be negative.

Some of the most useful feedback came from people on a Zulip forum. They pointed out where I'd overstated my claims, and they showed me how to prompt AI to critique my own work.

I keep every correction public in CORRECTIONS.md. If the code compiles, that only means the code compiles. I check whether the statement says what the paper claims.

I'm also learning the math myself, starting from the bottom. ChatGPT is coaching me through proofs and teaching me each concept in a way that I can grasp.