3 Scores Too Close to Call by Hashim
Hashim's entry into Varsity Tutor's November 2026 scholarship contest
- Rank: 184
- 0 Votes
3 Scores Too Close to Call by Hashim - November 2026 Scholarship Essay
The academic skill I most want to develop further is analysis, and I can point to the exact result that convinced me. This past summer I was an AI research intern at Fermi National Accelerator Laboratory, where I showed my mentors and other staff the scores from 3 benchmarks I had designed and graded. The scores for the 3 AI models I tested were too close together to say which model was better.
My project group and I spent about 4 weeks on a problem called hallucination, which is when an AI model confidently gives you an answer that it made up. We converted Fermilab documentation into a knowledge graph, so a model could answer through an exact query instead of generating a guess. One of my own parts was the testing. A benchmark is a test you give every model the same way so the scores can be compared. I wrote 3 original ones, ran 3 local models through them on Fermilab’s GPU farm, and graded what came back. The grading gave me 3 scores with almost no space between them. Scores that close can mean the 3 models really are about equal, or they can mean my benchmarks weren’t sharp enough to tell the models apart. Those are very different findings, and 3 nearly equal scores can’t tell you which one you’re looking at. I told my mentors what the numbers could support, which was that my benchmarks couldn’t separate the models. It’s a less satisfying thing to say out loud than naming a winner, and it’s the result I keep going back to.
In high school I founded Vitesse, an automotive apparel brand and media channel that grew to over 20 million organic views on Instagram. For 8 weeks I tracked the likes, comments, and saves on 64 posts by hand in a spreadsheet to find out which kinds of posts worked. I was sure the winner would be a clip of a gray Ferrari F430 with yellow stripes next to a lilac purple Porsche 911 GT3 RS outside Burdi European Boutique in Hinsdale. I had shot that clip well and color graded it, and it lost. The reel that won was a torque curve I filmed off a TV screen inside Dynojet Performance Tuning in Villa Park, from a Mercedes-AMG GT Black Series tuned to 875 horsepower. Across the 64 posts, technical reels like the torque curve beat the glossy videos by about 42% in saves and about 27% in comments, and I set my posting strategy from the spreadsheet.
The Vitesse analysis was the easy kind, because a 42% gap is big enough that a spreadsheet filled in by hand will find it. I think most of the differences worth finding in engineering are a lot smaller than 42%, and the Fermilab scores showed me what happens when a test meets a small difference. That’s the part of analysis I hope to develop further: deciding, before I collect anything, how far apart 2 results have to be before I can call one better.
I’m a freshman in mechanical engineering at the University of Illinois Urbana-Champaign, and I’m taking CHEM 103 lab this fall. I want to treat every lab as practice at asking how sure a measurement lets me be before I write the conclusion. In spring I plan to take PHYS 212, ME 200, and TAM 210, and I’ll carry the same question into each of them. I also plan to join Illini Motorsports, the Formula SAE team here, because a race car gets tested constantly and the difference between one setup and the next is usually small. Outside of class I run Elimenti, a web design agency I founded in March, and I send personalized cold emails to business owners every day. I want to track those emails the way I tracked the 64 posts, with the added step of writing down ahead of time what size of difference I would believe.
I’m in this degree because I want the engineering mindset, but really I want to go into entrepreneurship. My long-term goal is to start an engineering venture focused on mobility, one that builds physical systems that hold up under demanding conditions. The things I’m most interested in are vehicle dynamics, brake rotors, and the data that comes off a car while it’s being tested. When I guessed wrong at Vitesse, the cost was a reel that lost, and I could just post another one. A company that builds brake rotors doesn’t get that kind of second try, because by the time the data is proven wrong the part is already on somebody’s car. The decisions I want to be making at that company will depend on test results that are close, and somebody will have to say whether the gap between them is real. The next time I test 3 of anything, I want to have decided how close is too close before I see the scores.