Dylan Norquist
ML Researcher at DreamcatcherAI
Experience
DreamcatcherAI
Making LLMs fast. Research and engineering on LLM inference performance.
Direct Supply
Built internal AI tools for the elderly care industry: an agentic invoice validation engine (40 hrs/supplier saved), automated product tagging (cut manual labeling 93%), a product grouping system (160 hrs/week saved), and an integration testing framework.
Built a residual embedding finetuning model for product equivalency (+38% accuracy), used LLMs to expand abbreviations in unstructured product data, and wrote a menu PDF parser for recipes and nutrition.
Maintained a PostgreSQL/C#/React full-stack app and integrated internal AI for product equivalency, saving nursing homes thousands in procurement and rebate claims.
Education
Milwaukee School of Engineering
Awards & Research
SMEARGLE
Model-based speculative decoding that surpasses the previous SOTA (EAGLE3). Preliminary results presented at the Midwest Instruction and Computing Symposium (MICS 2026).
Proactive Urban Forestry Management
ML system for the City of Milwaukee that prioritizes tree pruning, shifting the city's forestry workflow from reactive to proactive. Mentored Xander Ede, Chukwuma Chukwuma-Ugwu, and Joshua Myers, who presented the work at MICS 2026.
Outside of Work
- Rock Climbing. I love indoor competitions and outdoor climbing (except lead climbing).
- Board & card games. Betrayal at House on the Hill is the favorite, with a close second in Poker.
- Puzzles. Rubik's cubes and whatever else I can find (sometimes puzzles find me).
- Minecraft. On and off since 2010, ~15 years deep. Recently got into MCSR and I'm quite bad at it.
- Challenges. Terraria Legendary Mode summoner-only, BTD6 CHIMPS, Borderlands 2 one-lifes, 100% Skyrim (haven't finished it yet), and more.