Skip to main content

Case Studies

Selected work from industry roles and my own ventures, labeled as such.

What this looks like in practice

The thread through my work: ML applied to the economics of expensive, constrained resources. Each story is labeled with the role I did it in, and claims are limited to what the public record supports.

NIL Valuation Platform

As founding engineer (contract) at ArcScore AI

The situation

ArcScore had a rough MVP and no production system: no pipeline, no infrastructure, and no defensible way to put a number on an athlete's NIL value.

What I did

I architected and built the platform end to end — a Next.js/React front end, a FastAPI back end, and a provider-agnostic Python data pipeline — and designed the core scoring and NIL-valuation methodology with two other modeling leads. I provisioned and now operate the production infrastructure: managed Postgres, a job queue, auth, secrets, CI/CD, and separate staging and production environments with observability on both.

The result

Live production platform. I own significant portions of the application, data, infrastructure, and engineering roadmap.

0 → 1
prototype to production

Capacity Planning Model

As an applied research intern at LinkedIn

The situation

The Applied Research team faced a confidential nine-figure hardware-allocation forecasting problem across LinkedIn's data centers.

What I did

I built a predictive model for it. The implementation is confidential, so the specifics stay with LinkedIn.

The result

13.39% improvement over the existing methodology.

13.39%
over prior methodology

Accelerating Physics Simulations

As a scientific ML research intern at AMD Research

The situation

Physics-based simulations at AMD were expensive to run, and an ML surrogate is only useful if you can tell when to stop trusting it.

What I did

I worked on a hybrid method that interleaves physics-based simulation with neural network training and inference, and on Bayesian and ensemble uncertainty-quantification methods that flag unreliable surrogate predictions on the fly.

The result

The method is public: I'm a named inventor on AMD's patent application US 2025/0005236 A1 and a co-author of a NeurIPS 2022 ML4PS workshop paper. The filing describes the hybrid workflow completing the simulation in less time than the baseline. It states no speedup factor, so neither do I.

Named inventor
US 2025/0005236 A1
Read the filing

Knowledge Retrieval System

As founder of TechnicallyFit, my own company

The situation

Critical know-how at TechnicallyFit was scattered across wikis, Slack threads, and people's heads.

What I did

I deployed a RAG system over the internal docs, so anyone could search and get real answers, not just links.

The result

One searchable source of truth instead of tribal knowledge, and less time spent reinventing the wheel.