Selected work

Research that survives contact with reality.

A selection of systems, studies, and open-source tools shaped by the same question: how do we make sophisticated methods reliable, legible, and useful?

01 / GENERATIVE AI

RAGLangChainOllamaLangSmith

Responsible GenAI, from architecture to evaluation

I design and teach applied generative and agentic AI systems that connect retrieval, orchestration, local model deployment, and monitoring—without losing sight of privacy or failure modes.

The work spans RAG pipelines, multi-step agentic workflows, local models with Ollama, LangChain orchestration, and LangSmith-based tracing. In parallel, I have helped shape a privacy-preserving synthetic transaction-data project using Gaussian mixture models and differentially private variants, balancing utility with client privacy.

ArchitectureRetrieval, orchestration, local inference, and observability

EvaluationBaselines, trace inspection, error analysis, and fit-for-use criteria

GuardrailsPrivacy, provenance, leakage, and human review

See the teaching practice →

02 / COMPUTATIONAL BIOLOGY

Interpretable MLGenomicsSHAPValidation

Reading the “genetic neighborhoods” of a poultry pathogen

Co-authored research tested whether genomic-island cassette organization could distinguish pathogenic from commensal Enterococcus cecorum lineages.

The study encoded biologically meaningful cassette architecture across 145 genomes and used locked, genome-grouped cross-validation to reduce leakage risk. The resulting representation outperformed simpler genomic-island burden baselines, while cross-project validation made the remaining deployment limits explicit.

145 genomes95 commensal and 50 pathogenic isolates

0.918 AUROCMean grouped cross-validation result for cassette-summary RF

Honest limitsProspective validation required before operational deployment

Read the Frontiers paper →

03 / NEURO-OMICS & AI

Multi-omicsExplainable AIGraph neural networksFoundation models

Connecting neuronal gene expression, multi-omics, and AI

Co-authored a 2026 review tracing computational neuro-omics from co-expression analysis to causal graph neural networks and cross-scale foundation models.

Published in Brain Informatics, the review examines Alzheimer’s disease, schizophrenia, and epilepsy, identifying data harmonization, spatial alignment, and causal interpretability as core barriers to translating multi-omic AI into reliable biological insight.

2026 reviewBrain Informatics 13(1):36

Cross-scale synthesisGene regulation, spatial transcriptomics, and neuroimaging

Translation challengesHarmonization, alignment, and causal interpretation

Read the review on PubMed →

04 / OPEN SOURCE

RCRANNeural networksTesting

nnR: making neural-network algebra executable

A CRAN package that turns mathematical constructions on neural networks into inspectable R operations and pedagogical examples.

nnR implements algebraic operations on neural networks and their approximations, connecting theoretical work on neural-network polynomials to code. The package includes documentation, a vignette, and test coverage designed to make abstract constructions easier to study and reproduce.

CRAN 0.1.0Published package with binaries and reference manual

PedagogicalDesigned as an understandable proof of concept

Research-linkedGrounded in peer-reviewed and preprint literature

Open nnR on CRAN →
Read the mathematical framework →

05 / SYSTEMS

RustBioinformaticsQuality control

RustQC: exploring leaner genomic quality control

An experimental Rust implementation of core sequencing quality-control ideas, built to explore portability, performance, and clearer deployment.

The project reworks parts of the familiar FastQC workflow in Rust and generates inspectable reports. It reflects a broader interest in pairing scientific correctness with systems-level efficiency.

View the repository →

06 / HEALTH ANALYTICS

Causal inferenceTime seriesShiny

From healthcare claims to decision-ready evidence

Applied causal, regression, Markov, and time-series methods to maternal-health and claims-volume questions in a health-insurance setting.

The work integrated claims with built-environment, social determinants, socioeconomic, Census, and commercial data. Reproducible analyses and R Shiny dashboards helped translate technical output into usable evidence for stakeholders.

07 / MEASUREMENT

Survey methodsPsychometricsProduct research

Measurement frameworks for people-centered products

Designed and applied statistical analyses for behavioral-product evaluation and children’s nutritional-needs research.

Projects included intervention comparisons using psychometric outcomes and complex-survey analysis using factor and structural models. The goal in each case was not just a result, but a defensible bridge from evidence to product or public-health decisions.

Working toolkit

Broad enough to connect the system. Deep enough to challenge it.

MODELS

Machine learning

PyTorch, TensorFlow, Keras, transformers, VAEs, diffusion models, BERT, LoRA, classical ML

SYSTEMS

Data + MLOps

Docker, MLflow, Nextflow, Bactopia, Airflow concepts, HPC, Slurm, Linux, Git

INFERENCE

Statistics + optimization

Causal inference, survey methods, glmnet, tidymodels, Pyomo, CVXPY, PySAL, Gurobi

Collaboration

Let’s move from interesting to useful.

If your project sits at the intersection of AI, evidence, and real-world decisions, I’d like to hear about it.

Start a conversation →