Field notes, methods, and references from our work with foundations, nonprofits, and mission-driven programs.
Medical AI is learned from a handful of places, then pointed at the people it never saw. The concentration holds at every level, and the fix is not more data. It is representation across the whole lifecycle.
Read the article →AI projects often fail for reasons that have little to do with model choice. Research across 44 medical AI studies points instead to data quality, infrastructure, interoperability, governance and workforce capacity.
Read →Health AI is only as representative as the data beneath it. When whole populations go uncounted, the models quietly fail the people who were never in the data — and it starts upstream, not in the algorithm.
Read →AI handles many languages well. But normally, patients in African clinics don't speak one language at a time — they braid English/French/Portuguese, a local tongue, and slang inside a single sentence, and most health AI, built for clean translation, breaks exactly there.
Read →Health and program exports hand you periods like "2026Q2" — strings, not dates. Every built-in time-intelligence function quietly fails on that. Here's the pattern for prior-period comparisons that actually works when your calendar is text.
Read →A decade ago, a Kenyan clinical information network published how it built trustworthy health data infrastructure under real constraints — no reliable internet, no budget for licenses, no EHR. The design principles behind that work have aged unusually well. Here's why they matter more now, not less.
Read →When a wrong number can change a clinical or funding decision, ordinary retrieval-augmented generation isn't enough. Supervised RAG — where the model routes to verified indicators and formulas instead of computing answers itself — is the pattern that holds up. A look at the approach, with a real public example.
Read →DHIS2 hands you data that is already broken apart by age, sex, and site — flattened into wide, awkward exports. Here is the star-schema pattern that turns those exports into a model your measures can actually trust.
Read →A practical framework for foundations and nonprofits deciding where artificial intelligence belongs in their portfolio — and where it does not.
Read →For mission-driven organizations working with sensitive populations, where data lives and who can read it matters more than any feature on a roadmap.
Read →Most RAG systems fail in the same predictable ways. A working architecture, the failure modes we have seen in production, and the design decisions that separate the two.
Read →A practical guide for nonprofit and foundation leaders evaluating data infrastructure proposals — the questions to ask, the answers to require, and the patterns that signal trouble.
Read →No articles match that search.
Practical data & AI methods for mission-driven organizations — the same rigor we bring to engagements, written to be used.
Twice a month at most. Unsubscribe any time via the link in every email.