Work
Built for real problems.
Open-data systems anyone can use, research on models that survive the field, assistants built in the open, and the methods behind enterprise platforms in production. One thread runs through all of it: from the raw table to the decision it informs, every number has to hold up.
How my work fits together.
01 Research
Research
Why models that excel in the lab fail in the field, and how to close the gap.
BroilerVision: computer vision for broiler houses
An overhead camera, a detector and a dashboard for the stockperson: it counts broilers with an honest range and flags any bird that stops moving.
Applied AI researchModelsongoingZimCropGuard: multi-crop disease detection under domain shift
Ongoing research on multi-crop disease detection under real African field conditions. Best Poster, Deep Learning IndabaX Zimbabwe 2026.
Tomato leaf disease detection: hybrid CNN-Transformer
My completed BSc Honours dissertation, deployed as a live Streamlit demo: a hybrid CNN-Transformer built to close the lab-to-field gap on budget phones.
02 AI systems
AI systems
Assistants and agents I build in the open: grounded in documents, guarded against misuse, and measured before anyone relies on them.
03 Open-data systems
Open-data systems
Public satellite, shipping and map data turned into series people can make decisions with, published with their failures.
Grid Atlas: electricity availability from night-time satellite imagery
Night-time electricity availability for 26 African cities from VIIRS imagery, each city measured only against its own history.
Logistics dataDatapreviewBorderFlow: corridor state for Southern Africa
Daily port throughput at Durban, Beira and Maputo from IMF PortWatch. The Beitbridge wait-time forecast is withheld until it can be scored.
Earth observationDatapreviewZimbabwe Crop Atlas: crop monitoring through the cloud season
Sentinel-2, Landsat 8 and 9 and Sentinel-1 from Digital Earth Africa, composited per pixel, with phenology that reports what it actually saw.
Geospatial data equityDatacompletedMapping equity: the flood crossings an open map cannot see
My Bias Bounty writeup: Texas low-water crossings the open map carries as ordinary roads, and an exact, reproducible build of the coverage gap score.
04 Enterprise delivery
Enterprise delivery
How I build at AI Creative Solutions, written as methods rather than client systems. Client work spans manufacturing, agriculture, healthcare and distribution.
AI agent orchestration
AI agents that answer questions on live business data within each user's permissions, with evidence for every figure and decisions left to people.
Data engineeringDataLive ERP integration and data pipelines
ERP and accounting data brought into lakes and warehouses without disturbing daily work, and checked before anyone relies on it.
Legacy modernisationPlatformsLegacy system modernisation
Replacing desktop databases and spreadsheet stacks with secure web platforms, with years of history brought across and every row accounted for.
Patterns I build in
- The source is the truth.Mirror the source system and reconcile against it. A disagreement is reported as a finding, never quietly fixed.
- Prove it before it goes live.Shadow runs compared column by column, and finance packs whose checks must tie.
- Unknown is not zero.A missing price or feed shows as unknown, in words, and never becomes a zero inside a total.
- Change and audit commit together.Every write records its audit event in the same database transaction, so the log cannot miss a change.
- People make the final call.Agents analyse, suggest and draft; decisions and approvals stay with a person, and AI actions are permission-checked and audited.