Postdoctoral Research Scientist | Northwestern University and University of Chicago, Pinto/Glaser Labs | Chicago, IL | Aug 2024 - present
- Develop ML and state-space models for high-dimensional sequential datasets spanning millions of data points across hundreds of complex behavioral and neural conditions, animals, sessions, regions, and task epochs.
- Build Python/PyTorch-compatible pipelines for preprocessing, feature construction, regression, decoding, GLM/state-space modeling, cross-validation, statistics, visualization, and failure-mode analysis.
- Test mechanistic hypotheses about hidden states, perturbations, and decision dynamics; translate ambiguous research questions into rigorous AI/ML analyses and reproducible computational outputs.
- Communicate assumptions, uncertainty, evaluation tradeoffs, and modeling decisions to computational, experimental, and engineering collaborators.
Swartz Foundation Postdoctoral Fellow / Postdoctoral Researcher | Princeton University, Pillow Lab | Princeton, NJ | Dec 2020 – Dec 2023
- Led ML research using latent state-space models, GLM-HMMs, and Bayesian inference to identify hidden strategies and transition drivers from millions of behavioral data points.
- Developed hundreds of lines of reusable modeling and analysis code for inference, validation, visualization, held-out evaluation, model comparison, robustness checks, and interpretability.
- Authored and presented a first-author Nature Communications study and collaborated with large multi-institutional teams on large-scale behavioral analysis.
- Extended and documented state-space modeling workflows in the SSM/GLM-HMM ecosystem, including input-driven transitions and observation models.
PhD Researcher, Neuroengineering and Machine Learning | University of Colorado / Anschutz Medical Campus | Denver/Aurora, CO | Sep 2016 - Dec 2020
- Developed GEMsort, an adaptive real-time multichannel spike-sorting framework using unsupervised clustering, dimensionality/data reduction, positional features, and hardware-aware constraints.
- Optimized algorithms for hundreds of thousands of neural voltage data points, including online cluster assignment and tradeoffs among accuracy, latency, memory, and computational cost.
- Designed and benchmarked efficient ML algorithms for large neural datasets, evaluating accuracy, robustness, computational efficiency, scalability, and real-time feasibility.
- Produced first-author Journal of Neural Engineering papers and a U.S. patent on adaptive real-time neural spike sorting.
AbbVie Internship, Biotech Company | Chicago, IL | Summer 2019
- Analyzed medical and engineering test data to support validation of wireless communication and electromechanical systems.
- Applied data analysis and ML-oriented quality-control thinking to evaluate device/test performance, detect inconsistencies, and support reliable experimental workflows.
- Validated equipment and workflows for Bluetooth
Sapienza University of Rome and Northumbria University at Newcastle, 2015 - 2016
Research Project: Visible light communication (VLC) systems
SELECTED PROJECTS & CODE
- PyTorch Transformer Interpretability - code: pytorch-AI-interpretability-transformer_ZM
Implemented a decoder-only transformer from scratch in PyTorch with causal self-attention, embeddings, residual streams, MLP blocks, training loop, and activation extraction; trained on non-ergodic hidden-process sequences and used PCA, effective dimensionality, and linear probes to test for interpretable belief-state representations. The project achieved R² > 0.989 for belief-state decoding and ~60% process-identity decoding versus 33% chance.
- Scientific Coding-Agent Benchmark - code: AI-coding-agent-benchmark_ZM
Built a reproducible benchmark for evaluating AI coding agents on real merged scientific-Python pull requests, with task generation, Docker-based sandboxed execution, automatic [0, 1] scoring, multi-model comparisons, log review, and failure analysis.
- Latent-State Modeling Pipeline for Nonstationary Decision-Making - code: glm-hmm_final
Developed end-to-end state-space/GLM-HMM workflows for preprocessing, model fitting, model selection, cross-validation, recovery analyses, and figure generation on large-scale behavioral datasets.
- Input-Driven State-Space Modeling Notebook - code: 2c-Input-Driven-Transitions-and-Observations-GLM-HMM.ipynb
Implemented and documented input-driven transition and observation models in the SSM/GLM-HMM ecosystem, showing how covariates modulate latent-state dynamics and observation probabilities.