Rohan Goel
San Francisco
I'm a data scientist at Roblox, on the Trust & Safety side. Most of my work is measurement: how well do automated systems actually catch harmful content, and how do you find that out without labeling the entire internet?
That turns out to be a statistics problem as much as a machine learning one. Where you spend a limited labeling budget, how you sample so the estimate holds up, what a benchmark number does and doesn't license you to claim. That's the part I find interesting. I'll stay vague about specifics here; most of the good detail is internal.
I studied data science and statistics at UC Berkeley. Outside of work I build small AI agents for things I actually use, train for treks I've talked myself into, and am circling the idea of starting something of my own.
Things I care about
- Evaluation
- Sampling design, variance reduction, and the gap between a benchmark score and a decision someone makes because of it.
- Agents
- Small, always-on systems that watch something on my behalf and are honest about what they don't know. Retrieval, tool use, and evals that catch regressions before I do.
- Mountains
- Long, slightly stupid group treks. Acatenango and Fuego in Guatemala last summer; Rinjani, Patagonia, and Kilimanjaro on the list. One a year is the goal.
- Skiing
- Winter weekends on snow, for the same reason as the treks.
- Training
- Lifting four days a week with zone 2 cardio around it, programmed around whatever trek is next.
- The 49ers
- Sundays from September through January, whatever else is going on.
- Fantasy football
- Where most of my side projects start. See Ask Madden and the draft simulator.
Elsewhere
What I've built is under projects, what I'm thinking through is under writing, and links has the rest: where to find me, and what's worth reading.