Keir Havel

Machine Learning Research Intern, Climate Informatics Lab · San Diego State University Research Foundation

San Diego, CA

• Built diagnostic tools to measure conservation-law violations in neural weather models (GraphCast, FourCastNet).
• Prototyped physics-constrained training methods for atmospheric conservation; showed available data was insufficient to measure conservation reliably, a negative result that redirected the lab's approach.

Data Scientist · Center for Sustainable Energy

Junior Data Scientist, Oct 2021 – Jun 2023Research Analyst Assistant, Feb 2019 – Oct 2021

San Diego, CA

• Integrated large language models (OpenAI, Anthropic APIs) into document review workflows, cutting processing time from 30 hours to about 1 hour while improving accuracy.
• Automated survey report generation, including statistical hypothesis testing and figure creation, reducing manual effort from 80 to 10 hours per cycle.
• Developed machine learning models for geographic vehicle traffic prediction using ensemble methods, with SHAP for interpretability.
• Built automated data ingestion (ETL) pipelines with data cleaning, enrichment, and anomaly detection.
• Improved budget forecasting by evaluating time-series models (ARIMA, additive models).

M.S. Applied Mathematics · San Diego State University

Concentration in Dynamical Systems. Coursework: Optimization, Bayesian Statistics.

Thesis: Compressibility and Multifractal Properties of Self-Compressing Neural Networks. Developed a regularization technique that substantially accelerated training in small-scale settings, and characterized how the speedup diminishes at scale.

B.S. Mathematical Sciences · University of California, Santa Barbara

Coursework: Data Mining (graduate level).

Fire: artificial agency

Experiments in agents that learn on their own: streaming deep RL, unsupervised skill discovery (METRA, CSF), cognitive map learners, and TD-JEPA.

Hippocampus models

A continuous-attractor grid-cell simulator and Vector-HaSH associative memory, extended with online Hebbian learning and metaplasticity tuned by evolutionary search.

Alternatives to backpropagation

Hebbian learning, predictive coding, forward-forward, PEPITA, reservoirs and evolutionary methods, benchmarked against backprop on sample efficiency and wall-clock time.

JEPA & Saccade JEPA

I-JEPA built from the paper before the official release, trained on ImageNet with one GPU, plus Saccade JEPA: an original variant that learns by predicting small eye movements.