Fig. 0 Machine Learning Engineer · Chicago, IL (Relocating to SF)
Innovating machines for more intelligent and secure experiences
Fig. 1 About
It started with FRC robotics on Team 2338, training neural nets to run on a Raspberry Pi strapped to a robot. That led me deeper into machine learning: teaching 200+ students as a UMN teaching assistant, training DQN autoscalers as a CGI intern, and along the way falling for security — how systems fail, how models get fooled, and what they leak about their data (that Modern Cryptography course stuck with me).
Today I'm a founding quant engineer at Crebit, building probabilistic pricing models that serve live FX quotes — full lifecycle work, from wrangling the data to training and evaluating honestly to shipping to production. Based in Chicago; building solutions to any problem that interests me.
Fig. 2 Work
Plate 01 Language models / efficient adaptation
A two-stage NLP pipeline that classifies character personality in film & TV scripts and rewrites it. LoRA fine-tuning cut trainable parameters by 90%, enabling Qwen2.5-32B on a single A100 with 4-bit quantization — 0.493 PAS / 0.838 BERTScore on personality transfer.
Plate 02 Forecasting / honest evaluation
Deep-learning forecasts of crop futures prices from NOAA and USDA climate data — temperature anomalies, precipitation indices, drought severity. An attention-based multi-stream LSTM beat PatchTST baselines: 5.9 vs 5.1 Sharpe over a 2-year out-of-sample period.
Plate 03 Sequence models / drafting
An MLP + RNN drafting agent for professional League of Legends — trained on webscraped match data, it reaches 70% similarity with professional drafts from Worlds 2025.
Plate 04 Agents / grounded explanation
The drafting system's next iteration: an agentic RAG layer where an LLM grounded in matchup data explains every pick in plain language. Explainability as a feature, not an afterthought.
Fig. 3 Experience
Founding Quant Engineer
Software Engineering Intern
Undergraduate TA
B.S. Computer Science
Fig. 4 Notes
Every few weeks I evaluate a new forecasting model for the crop futures project, and every time I re-learn the same lesson: the interesting bugs are in the backtest, not the model. These are the
Why this site exists, and how to add your own posts in about thirty seconds.
The stack behind this portfolio: an interactive neural-network canvas, hash identicons, markdown posts, and a Next.js static export on GitHub Pages.
Fig. 5 Contact
ML, quant systems, security, or something in between. Email is fastest.