Veejhay Roy — New Jersey, US

Researcher, developer & runner. Veejhay Roy

CS @ Georgia Tech ’30. I build things at the intersection of AI, mobile, and the outdoors — from an LLM spatial-reasoning benchmark to an iOS app used by 100+ athletes.

Selected work

YLookup × Encode — Hackathon · 1st Place

YLookup × Encode New

First place at the YLookup × Encode AI Hackathon in London. Team Gherkers built an agentic spreadsheet-editing harness around Qwen3.8-27B — inspect, write Python, recalculate, repair — scoring 88.25% on the 400-task SpreadsheetBench set.

All Work!

Handpicked highlights —
spanning the last few years.

Click any project to dive in.

RouteEval — LLM Research

RouteEval New

A benchmark for how well large language models reason about physical space — measured through the lens of generating real running routes. Built with MIT grad Roger Jin and submitted to ACL ARR.

Routecraft — Web App

Routecraft New

Turn a prompt into a running route. Routecraft generates routes with multiple models side by side so you can compare them — and hosts the RouteEval research blog.

SplitDecision — iOS App

SplitDecision

A coach-facing iOS app for tracking workout splits across cross-country and track teams. Built in Swift and used by 100+ athletes.

Fraud Detection — Machine Learning

Fraud Detection

Final project from Columbia University’s summer data-science program: a classifier combining logistic regression and a random forest, using SMOTE to balance a heavily skewed fraud dataset.

By the numbers

0+
Athletes using SplitDecision
0+
Route evaluations in RouteEval
0
LLMs benchmarked
’30
Class year, Georgia Tech

About

I’m a computer science student at Georgia Tech who likes building things end-to-end — research, software, and the occasional fast 5K.

Right now I’m researching how well large language models reason about physical space. That turned into RouteEval, a benchmark of 10,000+ running-route evaluations across 13 models, with a paper submitted to ACL ARR 2025.

In September 2026 my team took first place at the YLookup × Encode AI Hackathon in London, building an agent that edits real spreadsheets — 88% on SpreadsheetBench’s 400 tasks.

I also ship: SplitDecision, a coach-facing iOS split tracker used by 100+ athletes; a Columbia summer machine-learning project on fraud detection; and a HarvardX data-science certificate.

Outside of code, I run — cross-country and track, 400m through 5K, with all-county honors. Running taught me most of what I know about discipline, patience, and showing up.

Languages

  • Python
  • Swift
  • JavaScript
  • HTML
  • CSS

ML / Data

  • scikit-learn
  • pandas
  • NumPy
  • TensorFlow
  • SMOTE

Tools

  • Git
  • GitHub
  • Xcode
  • VS Code