Omar Pareja
Senior, B.S.E. Artificial IntelligenceUniversity of Pennsylvania, May 2027

Omar
Pareja

Now · October 2026What I'm building this month
SwiftDoor

AI-powered promotion, so organizers can reach the right communities and sell out their events.

VerifierGym

New environments for agents to train and be tested in.

01 / 04 · Founder · Dec 2025 to now

SwiftDoor

White-label ticketing and payments for parties, student organizations and live events.

Problem

Student organizers needed more control over their brand, ticket releases, and audience relationships than existing ticketing platforms offered.

Built

An end-to-end event platform with Stripe Connect payments and organizer payouts, customizable event pages, rule-based pricing and ticket releases, and QR check-in. A 3D audience map turns attendance data into an interactive network of groups and connections, with event-by-event replay and searchable audience insights.

Result
$1.2M+

In ticket sales processed across university events so far.

Visit swiftdoor.co ↗
How one episode is scoredVerifiers / PRIME-RL
Task→
Sandboxed site→
Agent in browser→
Grader→
Reward
02 / 04 · Builder · Agent training

VerifierGym

A training and testing ground for AI agents, with rewards you can trust.

Problem

To train AI agents with reinforcement learning, you need tasks that are graded automatically and correctly. Most agent benchmarks rely on fuzzy or AI-judged scoring.

Built

A realistic store where agents act through a real browser. Each task is graded by deterministic checks on the hidden database, so the score is exact and reproducible. 22 tasks, each automatically proven solvable and proven not solvable by doing nothing.

Hard part
8 of 8 fixed

An independent audit found 8 ways the grader could be fooled or could unfairly fail a correct agent. One let an agent pass with the wrong order total. I fixed all 8 and added a regression test for each.

Code is private. Happy to walk through it on a call.
03 / 04 · Co-Founder · Sep 2025 to Oct 2026

Caffio

A smart bottle that auto-dispenses caffeine to keep your energy steady between times you set.

Built

ESP32-S3 firmware over Bluetooth. A React Native app. A pharmacokinetic dosing model validated against published studies, with FDA-guideline safety limits.

Result
2,000+survey responses
50+beta users
Patentpending

Won at the M&T Innovation Fund and the Penn Wharton Innovation Fund Design Award.

Replaying history, twelve timesStand at a past date. Fit only on what came before. Forecast 60 days. Score it against what happened.
ActualForecast95% band28-day floorNaive baseline
Synthetic data. Not VideoAmp data.
04 / 04 · ML & Data Science Intern · Summer 2026

VideoAmp

I built the forecasts behind guaranteed-outcome ads: how many conversions VideoAmp can promise an advertiser before a campaign runs.

1.3B+rows in Snowflake
3advertisers
7daily series
12backtests per series

Experience

2023 to now
Omar Pareja
About

I like building real solutions, working directly with the people who use them, and owning the result.

Give me a messy problem and a real user, and I'll ship something that works.

B.S.E. Artificial Intelligence, University of Pennsylvania, 2027. Coursework: Graph Neural Networks (graduate), Natural Language Processing, Computer Vision, Algorithms.

Languages
EnglishFrenchArabicSpanish (intermediate)