Machine learning · Low-latency systems

Adam Thorne

Computer Science and Applied Mathematics at the University of Michigan

I like solving hard problems. That has led me to machine learning and low-latency systems—work where model quality, careful measurement, and efficient implementation all matter.

Selected work

Projects

Machine learning, quantitative research, and systems work.

Machine learning · Scientific computing

Physics-Informed Neural Networks for Financial PDEs

Inverse-PDE experiments for recovering local-volatility surfaces under controlled noise.

2–3 orderslower surface MSE

3noise levels tested

Python · PyTorch · PINNs · Neural ODEs · Autograd

View project

Machine learning · Quantitative research

Portfolio Research Toolkit

Shared data, validation, portfolio, and experiment contracts for comparing model-driven strategies.

4portfolio builders

3chronological split stages

Python · Polars · Parquet · cvxpy · QuantStats · MLflow

View project

C++20 · Low-latency systems

C++ Event Processing and Matching Engine

A deterministic event engine for measuring queue design, throughput, and tail latency.

4.2Mmessages per second

<35 μsp99 latency

C++20 · CMake · CTest · SPSC/MPSC queues · Sanitizers

View project

C++20 · Query execution

Vectorized Columnar Analytics Engine

A typed columnar engine for scan, filter, projection, and hash aggregation.

39.3×scan p50 speedup

5.24×group-by p50 speedup

C++20 · CMake · CTest · DuckDB · Python benchmarks

View project

Work experience

Experience

Machine-learning research, evaluation, and tooling.

Lightshift Energy

Machine Learning Analyst Intern

Summer 2026

Automated model training, HPO, scoring, and comparison across roughly 1,800 configurations and 620 HPO trials over about 800 GPU-hours.
Reached 1.88% MPE, 63.4% capture@1, and 100% capture@4 across 155,952 forecasts with roughly 4× lower MAE than a baseline.
Raised capture@1 by 9.2 percentage points versus production while holding MAE within 0.02 MW.
Benchmarked spatial autoencoder weather representations against legacy and competitor encoders, reaching 100% capture@3/@4 and improving capture@3 by 3.03 percentage points.
Reduced model-selection risk with production-faithful scoring harnesses, reproducing deployed metrics within noise and resolving peak-set and capture-metric discrepancies.
Profiled the training data path to reduce epoch time from 30.3 seconds to 8.6 seconds and a projected 12+ hour pipeline stall to 15–20 minutes.

Verazoi

Founding Machine Learning Engineer and Researcher

February 2026 — present

Built multimodal event-forecasting pipelines across 44,348 windows and 177,392 horizon targets.
Reached 0.853 mean test AUPRC and 79.1% F1 at 60 minutes.
Used time-aware splits, horizon-specific labels, threshold policies, and tracked evaluations.

Machine Learning Student Network

Project Manager and Machine Learning Engineer

October 2025 — present

Built shared data, prediction, portfolio, backtest, reporting, and MLflow contracts for a quantitative research team.
Helped researchers debug leakage, metric choice, baseline design, and portfolio-level interpretation.

How I work

Define the problem. Build the simplest serious solution. Measure it honestly.