Software
I build software and data infrastructure when a methodological contribution must remain reproducible, inspectable, and usable beyond a single study. The systems below support two parts of my research program: validating cause-of-death methods and making mortality reference data easier to discover, evaluate, and reuse.
Reference Data Archive software ecosystem
JuliaPythonRSQLAPIsData engineering
- Problem
- Mortality reference datasets are difficult to discover, document, integrate, quality-check, and reuse consistently across institutions.
- Deployment
- Operational software supporting the WHO-hosted public catalogue and secure analytics environment for mortality reference data.
- What it enables
- Ingest and integrate mortality reference data from research sites and partner organizations.
- Harmonize variables and metadata, validate incoming data, and automate quality-control workflows.
- Generate and publish DDI metadata to a NADA catalogue and provide programmatic access through user-facing APIs.
- Support platform administration, controlled data access, user management, and secure analysis workflows across collaborating institutions.
- Related outputs
- WHO platform · RDA GitHub organization · Consulting engagement
MultimodalVA
PythonNLPMachine learningMultimodal learningLLMs
- Problem
- Automated verbal-autopsy methods commonly discard free-text narratives or evaluate narrative and structured symptom data in separate workflows.
- Deployment
- Open-source, MIT-licensed Python package archived at Zenodo.
- What it enables
- Prepare narrative and structured verbal-autopsy records for reproducible cause-of-death modeling.
- Train and compare narrative-only, indicator-only, and multimodal classifiers under consistent data splits and metrics.
- Run data-fusion, feature-fusion, soft-voting, and stacking models with hyperparameter optimization.
- Run prompt-based LLM pipelines with open-weight models loaded from Hugging Face or served locally through Ollama; closed-source API models are not currently supported.
- Evaluate and visualize model performance under consistent individual- and population-level metrics.
- Related outputs
- GitHub · Zenodo release · medRxiv study preprint · Dissertation preprint · Software overview
Methods & Tools
- Programming & statistical computing
- R · Python · Julia · SQL · Stata · JAGS · NIMBLE
- Machine learning & language models
- Hugging Face · Optuna · pretrained language models · multimodal learning · prompt-based inference · Ollama
- Data systems & delivery
- Data integration and harmonization · metadata and DDI · REST APIs · ODK · NADA · Azure · Kubeflow · Snowflake · AWS
