How Did YIMO Build An AI Proctoring System To Host 3,000 Math Competitors For Free?
Most students who love math competitions find out early whether they are allowed to enter one. The answer usually has nothing to do with mathematics. It comes down to a registration fee, a travel budget, or whether a school decides to sign its students up at all.
The Youth International Mathematical Olympiad, known as YIMO, was built so that none of those questions come up. It was co-founded in February 2026 by Hyunjun Yi and a partner based in New Jersey, who were already running two small STEM nonprofits, NXT Horizon and STEMise, and had both noticed the same group of students going unserved. The competition is free, fully online, and open to anyone from roughly age nine through college. Within months, it drew 3,000 participants across 38 countries.
The Name Came From the Mission
The Y in YIMO was not decided after the fact. The co-founders wrote the mission statement first and named the organization from it. The target was young people who wanted to do competition math and could not, for reasons that were economic, geographic, or simply structural. That founding principle shapes the decisions that follow it. Papers are translated so that language is not the thing standing between a student and a problem set. Beginner and Advanced divisions run separately, so a newer competitor is measured against peers rather than against students who have been training for years.
The team behind it is not designing for a community it has only read about. Its 45 staff members, spread across 21 countries, include former olympiad competitors with USA(J)MO, MOP, USAPhO, USACO, and IOAI backgrounds, and Hyunjun holds a perfect score on the AMC 12 and has placed in other competitions. Most are returning to a circuit they came up through, this time from the other side of it.
Free and borderless solves one problem and creates another. If anyone can compete from any room in the world, with no invigilator at the front and no school to vouch for them, why should the results mean anything? Hyunjun had competed in online math competitions before co-founding one, and the recurring flaw was always the same. Dishonesty was common and almost impossible to catch. Once YIMO began reaching thousands of participants, watching everyone closely enough stopped being something a human team could do alone.
So YIMO proctors every round twice. Live human proctors watch each contestant over video, camera on and desk in frame. An AI system watches the same feed. It looks for prohibited items and situations, including phones, headphones, calculators, and a second person in the room, and it tracks how long each occurs. From that, it generates a risk score. The higher the score, the stronger the case that something is worth a closer look.
Inside the AI System
The hardest part was not the concept but the training data. Off-the-shelf detection was never going to be accurate enough, because not every phone looks like the same phone and not every instance of cheating looks the same. Hyunjun, who led the system’s technical development, wrote code and built the dataset himself. Staff joined mock Zoom sessions and behaved like dishonest competitors, working on a phone below the desk, wearing headphones, running an AI tool on the side, so the model could learn the range of what it would encounter. A collaborator with a master’s in data science and artificial intelligence from Harvard, who works as a research scientist, supported the modeling work.
One design decision was deliberate and early. The system sees the camera and never the screen. The camera captures the physical setting, which is where the evidence is, and the narrower scope keeps the tool closer to detection than to surveillance. Participants sign a consent form stating that footage is used only for proctoring and is deleted once the process is complete.
Why a Human Always Makes the Call, Not the AI
The risk score is a report, not a verdict. Human proctors are already in the room. When the model raises something, a person reviews the footage and decides what it means. The division of labor is simple: the AI reports and humans judge. It is an easy line to state and a harder one to hold as volume grows, because the appeal of automation is strongest precisely when the caseload is largest.
The case that proved the point came from an unexpected direction. The model began flagging translated versions of the paper, reading an accessibility feature as a prohibited resource. A system with final authority would have penalized students for using the exact provision built to include them. Human reviewers caught the pattern and corrected it, and the translation program stayed in place.
Detection accuracy holds up in smaller groups and degrades as rooms fill, so thousands of participants have to be split across many Zoom rooms, and every additional room needs another proctor. The team names capacity without loss of accuracy as the hardest fairness problem YIMO faces, and does not claim to have solved it. That willingness to leave a problem open is the same instinct that keeps a person on every consequential decision.
What It Costs to Stay Free
None of this works without funding, and the funding began with cold outreach. The first sponsor was Art of Problem Solving, a well-known name in competition math, which contributed after YIMO emailed the organization without an introduction. PiMath, a math education organization run by a Stanford-affiliated professor, followed, and Hudson River Trading came later. Between them, they fund a prize pool worth $2,500, awarded to students who paid nothing to compete for it. The staff team runs on Discord, with weekly all-staff meetings and biweekly directors’ meetings, a structure the team describes in terms of transparency rather than oversight.
For Hyunjun, who intends to study artificial intelligence, the proctoring system is also a first argument about how the technology should be used, and it is a restrained one. Give the model a narrow task. Point it at the desk and not the screen. Delete the footage. Put a person in front of every decision that affects a student. Those are the choices of a team that has thought as carefully about what the technology should not be asked to do as about what it can.
YIMO is set out to build a competition for students who had been priced out of the ones that already existed. Making it trustworthy turned out to be the harder engineering problem, and the more interesting one, that Hyunjun and his co-founder are successfully and rigorously trying to solve, and have done so to a great extent.
