Independent research / PyTorch
Physics-informed machine learning
Wrote a 29-page paper on local-volatility calibration from noisy option prices, comparing inverse PINN and neural-adjoint solvers on error, smoothness, and runtime.
Hi, I'm
Adam
Thorne
I'm a computer science and applied math student at Michigan, interested in how models represent data and predict what comes next.
Take a look aroundAnn Arbor, Michigan
Draw a number
It guesses as you go ONE DIGIT, 0–9
Inside the encoder32 values
Play with the values
Draw a number to explore its values
Pick any square, then move the slider to see what the model decodes on the right
What am I looking at?
Each square shows one value in the model's shared representation of your drawing. Darker squares are more active. Select a square to change its value, or read the full vector below. Both predictions use this same state. Edited values are decoded by the model and may no longer describe your original drawing. The squares aren't assigned meanings like “a loop” or “a line”.
MY GUESS 7 Reading your handwriting Other guesses — Possible finish
A closer look at the work
Research and things I've built
PyTorch / ResNet-50 / LSTM
Multimodal corn futures forecasting
Fused ResNet-50 satellite embeddings, weather, seasonality, and futures prices in an LSTM to forecast corn price changes; used chronological splits and train-only normalization.
Current work / Triton / LightGBM
ML-guided GPU kernel tuning
I'm working on a LightGBM selector that maps LLM tensor shapes and dtype to Triton tile sizes and warp counts, with latency comparisons against autotuning that include selection overhead.
Where I've been working
Roles and research
Lightshift Energy
Machine Learning Engineer (Part-Time)
Sep 2026 — Present
- Applied machine learning to improve an automated optimization workflow.
Machine Learning Analyst Intern
Jun 2026 — Aug 2026
- Developed and evaluated PyTorch models for forecasting.
- Rebuilt a model-comparison library with Polars and parallel execution to reduce analysis time and repeated data reads.
- Used agentic coding to automate model configurations and GPU experiment sweeps.
University of MichiganLei Ying Research Group
Undergraduate Researcher
Sep 2026 — Present
- Ran VLA and RL experiments: fine-tuned SmolVLA for 5,000 updates, replay-verified 20 Isaac rollouts, and audited 1,152 poker-policy evaluations.
Verazoi
Founding Machine Learning Engineer and Researcher
Feb 2026 — Present
- Led technical discussions during pitches, translating model design, validation risks, and user impact into a clear product case that helped secure $2,000 in funding.
- Built a multimodal prediction pipeline across 44,348 windows from 45 participants, reaching 79.1% F1 for one-hour glucose-event prediction with chronological tests and MLflow.
Machine Learning Student Network
Project Manager & Machine Learning Engineer
Oct 2025 — Aug 2026
- Led technical lectures that helped members progress from linear regression to deep learning within two weeks.
- Built a shared Python ML framework that cut model development and testing time by 90% across project teams.
Oxygen Coaching Group
Software Engineering Intern
Jul 2025 — Sep 2025
- Built a TypeScript/Electron desktop app to automate employee wage calculations.
A bit about me
I'm studying computer science and applied mathematics at the University of Michigan, after two years at UC Davis.
I like the part of ML where the model meets a messy problem: getting the data into shape, deciding what a fair comparison looks like, and figuring out what the results actually tell us.
at@adamthorne.com