I build AI evaluation systems and multi-agent tools, mostly in Python and TypeScript. Industrial Engineering + AI at Northwestern.
Before software I ran a 65-person newsroom and won a national debate title. Strange background for an engineer, and the most useful one I could have asked for: I still think the hard part of any system is explaining why it did what it did.
AI/ML, the main event: LLM evaluation, RAG, LangGraph and multi-agent orchestration, embeddings and clustering
Web
Data and infra
Most of what I build now lives in private startup and research repos. These are the public ones. Four of them sit under teammates' accounts because that's where we started them. GitHub lists me as the top contributor by commit count on each of those four.
tr-benchmarking (Thomson Reuters, CS+Law Innovation Lab) Reasoning-level evaluation for legal LLMs. Forces IRAC-structured output, embeds it, then clusters with UMAP and HDBSCAN to find where models diverge doctrinally. Adversarial poisoning tests prove the clustering actually isolates bad reasoning. Cut expert review time by 90%.
Architec (WildHacks, 3rd of 70 + track win) Energy audits that normally cost $50k, done in 10 minutes. Gemini reads your utility bills, weather-normalized regression splits heating from cooling from baseload, and a Three.js digital twin shows where the building leaks heat.
LaborLens (Uncommon Hacks, Social Impact finalist) Seven-agent LangGraph swarm that traces labor-exploitation risk through a supply chain, pulling from news, sanctions lists, court records, and country risk indices, and citing every finding.
GreenChain (HackPrinceton, 3 sponsor tracks) Ranks manufacturers by real environmental cost. An agent swarm on Dedalus Containers finds them, XGBoost quantile regression over USEEIO and Ember Climate data scores them.
Arrival (ALI Builds, 1st place) A credit profile for the 45 million people the US credit system can't see. Turns foreign rent history and bank statements into a portable trust profile, issued as an Apple Wallet pass.
Meritus Labs. LLM evaluation infrastructure, a quality layer for teams putting AI in front of customers.
Northwestern Medicine (FORGE). Weakly supervised models predicting MSI from colorectal whole-slide histology, in the inaugural research cohort.
WingRep. GTM engineering, building AI outbound pipelines.
I'm looking for Winter and Summer 2027 internships in AI/ML, evaluation, and data systems. Always happy to talk about agent architectures, or about why your eval suite is lying to you.
