Udit Ranasaria

Udit Ranasaria

Director of Research, SumerSports

I'm a full-stack deep learning researcher at SumerSports, where I build foundation models for football. My tokens are player trajectories instead of text or patches. I lead the research team and set our technical direction while spending most of my week inside the stack: cleaning and augmenting data, designing architectures and training recipes, and building evals. The models run in production behind analytics for NFL teams.

I studied computer engineering at Carnegie Mellon and started out as a database systems engineer at Microsoft. I taught myself data science for the NFL Big Data Bowl, which led to a research job with the Cleveland Browns. The part of that job I loved was unlocking insights from large, messy tracking data. I joined SumerSports in 2022 to do that at larger model scale, and taught myself deep learning there. We later built and open-sourced SportsTrackingTransformer.

I also teach. I've spoken several times at the Carnegie Mellon Sports Analytics Conference, I help advise a team of student researchers at UNC Charlotte extending my transformer work, and in September 2026 I'll lead a student transformer workshop at CASSIS, a sports analytics conference in Vancouver.

Udit Ranasaria

Research, open source & writing

Public Work

At SumerSports I architected our foundational deep learning model for football, a transformer that learns player movement directly from raw tracking and can generate synthetic plays. I open-source what's useful and write when I can.

SportsTrackingTransformer architecture diagram: tracking data into a transformer player encoder

SportsTrackingTransformer

open source

Attention Is All You Need, for Sports Tracking Data

The first end-to-end transformer for NFL player tracking that treats all 22 players as a permutation-invariant, unordered set. No hand-ordered features, no positional templates, just raw tracking in. It beats the Big Data Bowl-winning “Zoo” baseline by roughly 20% on tackle-location prediction, and pulls further ahead as plays develop.

Where it went:

  • 2025 Big Data Bowl — overall winner. “Exposing Coverage Tells in the Presnap” cited the framework as what led them to a transformer architecture, and adopted its training methodology.
  • NFL Next Gen Stats. Productionized the winning transformer as “Coverage ID” on Amazon Prime Vision broadcasts.
  • 2026 Big Data Bowl — finalists. UW–Madison’s entry built its transformer on the approach and used SumerSports’ open data supplement.

Code Paper

Open-source data release

Big Data Bowl 2026

Model-generated coverage data + transformer tutorials

Frame-level coverage schemes and player-level assignments, generated by our models and published openly for the 2026 Big Data Bowl, alongside transformer starter notebooks. We put it out so anyone could build on the same foundation we use internally. Big Data Bowl finalists did.

Data Tutorial Write-up

Passing Value in Expectation: a pass arc over a field with player influence probabilities

PaVE the Way for NFL Passing Analytics

Honorable Mention · 2021

Passing Value in Expectation — NFL Big Data Bowl

My entry point into football analytics. On a four-person team, I built the pass-completion simulations from physics-based projections. The project earned an Honorable Mention and led directly to a season of research with the Cleveland Browns.

Notebook Code

Simulated player-tracking on a basketball court degraded by sensor timing noise

Low-Cost Sports Real-Time Location System

Senior Capstone, CMU ECE · 2020

Inexpensive real-time player tracking from low-cost hardware

Where tracking data first hooked me. As design lead for my senior ECE capstone, our team built an affordable ultra-wideband system to locate players on a basketball court in real time. I designed the time-synchronization algorithm and programmed the simulators we used to measure how sensor timing noise and anchor placement degrade tracking accuracy.

Video

Teaching, talks & mentoring

Teaching

Teaching is where a lot of this started, and I keep finding my way back to it.

The CMU CS Academy penguin

CMU CS Academy

Pedagogy Content Lead · 2017–2020

I led the pedagogy content team for several semesters of Carnegie Mellon's free CS curriculum, now used by high-school classrooms worldwide. I designed programming exercises, guided the content creators building them, and mined usage data to find which lessons were landing and which needed a rewrite.

Site Team

Intro to Programming, CMU SCS

Teaching Assistant · 2017–2020

Ran recitations and office hours, graded exams, and advised term projects, with a hand in curriculum revisions for the course.

Advising student research, UNC Charlotte

2026

Helping advise a team of student researchers who are extending my transformer work on sports tracking data.

Talks & workshops