Bayesian, geospatial and AI decision systems for immunisation, HIV and epidemic response, built to research standards and handed over to the teams who use them.
Most health data describes the past. We build the part that tells a programme what to do next, held to research standards: every estimate carries its uncertainty, every model is tested against data it never saw, and every platform is handed over.
15+
years directing evidence toward health resource decisions across sub-Saharan Africa
8,871
health facilities tracked on 27 harmonised indicators for UNICEF's AHEAD programme
497,620
people on HIV treatment in our twelve-year survival cohort in Nigeria
87,946
programme records behind our HIV testing-yield forecasts across 29 countries
Commissioned by NPHCDA, UNICEF, Gavi, USAID, PEPFAR and the Henry Jackson Foundation, with methods peer reviewed before they are deployed.
A marked finger shows one child was reached.
A health worker marks the finger of a vaccinated child.
Our work is finding the children without one.
The old city of Kano from the air, 1959.
Routine reports, household surveys, programme records and spatial surfaces arrive with different periods, boundaries and gaps. We reconcile them to one geography, then model them in open, reproducible code that the client keeps.
Drag the globe to turn it.
20%
lower four-quarter forecast error than a persistence baseline, in EvidenceLab Program Forecast Studio
Bayesian and Prophet time-series models forecast coverage, dropout and programme indicators with prediction intervals, and flag where delivery falls below an area's own recent level.
A flag is a decline tripwire, not a coverage figure, and we say so on every screen. Our model cards also say where a model does not beat its baseline: one quarter ahead, this one does not.
37
states, counting the Federal Capital Territory, in our Bayesian projections of paediatric HIV and unmet treatment need, 2025 to 2030
Hierarchical models let states and districts borrow strength from one another, so sparse and uneven surveys still yield stable estimates with honest intervals.
Our Bayesian estimates of HIV prevalence and burden in Nigeria, published in EClinicalMedicine, were used by the PEPFAR interagency team and the National Agency for the Control of AIDS.
774
local government areas in Nigeria's national zero-dose model, built for NPHCDA; findings forthcoming
Small-area estimation turns national surveys into estimates for every district: constrained to agree with the survey at state level, and smoothed spatially so neighbouring areas borrow strength.
Each area gets its own credible interval and a probability of being a priority, so planners can see where the evidence is firm and where it is thin.
93%
classification accuracy for HIV case-finding models across five countries, Henry Jackson Foundation
Spatial scan statistics, Getis-Ord Gi* hotspots, empirical Bayesian kriging and geographically weighted regression show where a problem clusters and which barrier dominates where.
We have applied them to key-population HIV data in Nigeria (JMIR Public Health and Surveillance, 2021), to regional HIV drivers across Tanzania's 31 regions, to COVID-19 in sub-Saharan Africa, and to paediatric HIV case finding in Kenya.
497,620
people on HIV treatment in a twelve-year retention cohort in Nigeria
Kaplan-Meier, Cox and competing-risks models answer when, not only whether: time to loss from care, time to an event, time since a pathogen was last detected.
Unsupervised clustering of 11,589 records reported as lost to follow-up found six distinct profiles within a single programme label (HIV: Current Research, 2025).
7
languages and more: English, French, Hausa, Yoruba, Igbo, Pidgin, Swahili
The Zero-dose Analytics and Risk Assistant answers plain-language questions using only the live model outputs, cites uploaded documents by page, and is instructed never to invent a number.
It sits inside a no-code platform that re-runs every model on newly uploaded data, owned and operated by the agency that uses it.
A programme does not need a model. It needs an answer it can act on this quarter, with the uncertainty stated and the working shown. Each question below is one our models answer for the teams who commissioned them.
For the National Primary Health Care Development Agency, with UNICEF and Gavi, we built Nigeria's first national predictive model of zero-dose children, for every local government area, and handed it over as a platform the agency runs.
The findings will be released shortly. The full study, its methods and an interactive map will appear here at that point.
NPHCDA with UNICEF and Gavi, January to July 2026. A national predictive model of zero-dose children for every local government area, handed over as a platform the agency runs. Findings will be released shortly.
About the engagement8,871
health facilities tracked on 27 harmonised indicators, from DHIS2 extraction to decision dashboards
An end-to-end pipeline from DHIS2 to multilevel models and dashboards across maternal and child health, HIV, immunisation and nutrition, for the AHEAD programme, with multilevel models that show whether a gap sits at the facility, the district or the national level.
All engagements15%
more service uptake in Tanzania after geospatial targeting redirected effort to high-yield areas
Quality metrics, analysis and scientific reporting for the Walter Reed Army Institute of Research environment across five countries, 2022 to 2025. In Tanzania, Lasso regression and geographically weighted regression across all 31 regions showed that HIV drivers differ by region, so prevention and case finding should too.
All engagements| Year | Work | Where | Recognition |
|---|---|---|---|
| 2026 | AlignInsight: detecting deceptive alignment in healthcare AI | medRxiv | Preprint |
| Paediatric HIV hotspots in Kenya: machine learning and geostatistics | medRxiv, 2026 | Preprint | |
| 2025 | Forecasting inequity: paediatric HIV treatment gaps in Nigeria | Preprints with The Lancet, SSRN | Preprint |
| Loss to follow-up on ART: cluster analysis with unsupervised machine learning | HIV: Current Research 10:425 | Peer reviewed | |
| 2024 | HIV Vaccine Awareness Day symposium | Walter Reed Army Institute of Research | Best oral presentation |
| 25th International AIDS Conference, Munich | International AIDS Society | Presentation award | |
| 2023 | Bayesian estimation of HIV prevalence and burden in Nigeria | EClinicalMedicine 62:102098 | Used by PEPFAR and NACA |
| New Investigator Scholarship | CROI 2023, Seattle | Scholarship | |
| 2022 | Case finding through community-led index partner testing | EClinicalMedicine 43:101265 | Peer reviewed |
| Survival analysis of retention on antiretroviral therapy | CROI 2022 | Presentation | |
| U.S. Embassy Eagle Awards (2017, 2019, 2021, 2022) | U.S. Embassy Nigeria | Awards | |
| 2021 | Geographical disparities in HIV among key populations | JMIR Public Health and Surveillance | Peer reviewed |
| COVID-19 infection fatality rate for 48 African countries | BioMed 1:63–79 | Peer reviewed | |
| 2020 | Swiss National Science Foundation grant 320030_192452, CHF 1,008,000 | University of Geneva | Research grant |
Silver Spring, MarylandQuantium Insights LLC
Makurdi, Benue StateQuantium Insights Nig Limited, Nigerian affiliate
EmailDirect to the principal
Dr Amobi Andrew Onovo founded Quantium Insights after fifteen years directing evidence toward resource decisions across sub-Saharan Africa: Principal Investigator of Nigeria's national zero-dose modelling programme, Technical Director at the Henry Jackson Foundation in the Walter Reed research environment, and Activity Manager for a $15 million USAID-PEPFAR agreement.
He holds a PhD in global health (Bayesian and spatial modelling) from the University of Geneva, an MPH in epidemiology and a postgraduate diploma in data science from UT Austin, and he has trained government and partner analysts in five countries. More about the practice.
Have a question your data cannot answer yet? Tell us the decision you need to make.
We take on a small number of engagements at a time, and we say early and plainly whether the available data can support the question.