Taking engagements for Q3/Q4 2026 through Samplics LLC: sample design review, weighting and variance estimation, small-area estimation pilots, bid support, and training. English or French, remote or on-site.
Recently completed: NBS Tanzania capacity assessment for SADC, funded by the World Bank; INSBU Burundi small-area estimation training (2 weeks); FeeLoST DHIS2-integrated platform for Ethiopia's Ministry of Health.
Methods & research
Building svy-sae, the small-area-estimation engine of the svy ecosystem (JAX): Fay–Herriot and unit-level models that fuse survey with administrative, census, and remote-sensing data, with proper MSE.
Writing the fusion-SAE methods-and-software paper as the flagship: scalable small-area estimation from multi-source data, shipped in production software.
Also svy-causal: survey-weighted IPTW and doubly robust estimation, validated against NHANES — being written up as a methods paper. A bounded methods probe, not a change of direction.
Platform
svyLab is at 327 passing tests and approaching public preview: the analytics platform built so that survey-correct estimation is reproducible and auditable by construction, with AI-assisted analysis that always shows its work.
Frontend in Astro + Svelte 5; backend in Python (Litestar) + PostgreSQL + DuckDB + Redis; AI layer via svy-agents.
Career
The consulting practice is the main line of work. I also consider senior roles in survey methodology, small-area estimation, and population-health measurement — full-time or contract — at statistics offices, health agencies, foundations, and research institutes.
Either way, most interested where the work is producing trustworthy local estimates from fused survey, administrative, and geospatial data, and where rigorous methods and reproducible software both matter.
Reading & thinking about
Fusing survey, administrative, census, and geospatial data for reliable subnational estimates, and the uncertainty quantification that keeps them defensible.
Transportability and survey-weighted causal inference: extending design-based methods to populations a sample was not drawn from.
How AI-assisted analytics can get the statistics right by construction, rather than producing confident wrong numbers.
If something here resonates with what you're working on,
get in touch
.