Latest updates are temporarily unavailable. Showing the research overview.
We connect explainable AI research, healthcare innovation and practical assurance to help institutions build confidence in consequential AI systems.
CognitiveAI Labs brings together AI research and applied systems development, with a focus on emerging markets and high-impact decisions.
Our work spans multimodal healthcare AI, quantitative explainability, trustworthy autonomous agents and evidence-based AI assurance. We develop the methods and tools needed to examine how systems perform, where they may fail and how people can retain meaningful oversight.
Healthcare innovation and AI assurance, connected by a shared commitment to explainability, evidence and human oversight.
Our healthcare platform brings applied AI into care workflows. Its research foundation includes multimodal tuberculosis screening and explainable predictions.
The supporting system for the CognitiveAI Assurance Framework (CAAF), connecting controls, evidence, assessor review and assurance decisions.
Our research connects model performance with explanation quality, operational safeguards and accountable decisions.
Research combining chest X-ray imaging and clinical metadata through late fusion, with quadrant-based activation analysis and perturbation tests to evaluate explanations.
Integrating imaging and structured metadata to study how complementary evidence influences screening predictions.
Moving beyond visual heatmaps to examine activation patterns, stability and sensitivity through measurable tests.
A research direction exploring runtime evaluation, unsafe-action detection and meaningful human oversight for autonomous agents.
Combining image-based and clinical evidence at the decision level to study tuberculosis screening.
Two independent predictions. One combined risk estimate.
The model generates a tuberculosis probability from chest X-ray images using a convolutional neural network (CNN), and a separate probability from clinical metadata using a Random Forest model. A late-fusion Random Forest meta-classifier combines the two probability outputs.
The model and its companion Grad-CAM module form a research workflow for examining classification, calibration and explanation quality.
Intelliscan TBCNN model
Image-based TB probability
Random Forest model
Metadata-based TB probability
Combines the matched probability outputs into a final TB risk estimate.
Independent image and metadata training, probability generation and late-fusion evaluation, with confusion matrices and classification metrics.
Reliability curves, expected and maximum calibration error, threshold analysis and decision curves support scrutiny of model outputs.
The companion module provides Grad-CAM overlays, quadrant-based activation analysis, perturbation evaluation and exported results.
Research system · This website introduces the model; live inference is not connected.
Safety, fairness and accountability claims require measurable, reviewable evidence.
A failed critical safeguard cannot be compensated for by a strong aggregate score.
Monitoring, intervention and reassessment must continue as systems and their operating conditions change.
Connecting technical research with institutions, policy perspectives and the communities AI systems affect.
Convened by Furaha Marwa, the roundtable brought together perspectives across technology, policy, law, finance, digital infrastructure, sustainability and civil society to examine practical, evidence-based assurance.
Participants agreed to establish a coalition of partners and experts and to share the framework’s controls for review, with each assurance domain assessed for practicality.
Multilingual populations, demographic and geographic variation, underserved communities, infrastructure and data sovereignty must shape assessment. Contextualisation should strengthen relevance without lowering assurance expectations.
The roundtable was conducted under the Chatham House Rule. Outcomes reflect collective themes and agreed next steps; they do not imply that every participant endorsed every formulation or that the framework is a completed certification regime.
Our international engagement includes participation in UN Digital Cooperation Day and convening the September 2026 expert roundtable during UN General Assembly High-Level Week.
New research and engagement updates will appear here as they are published.
Photos from UN engagement, the roundtable and research will appear here when published.
Connect with us about research partnerships, the TB Late Fusion Model, framework review, institutional pilots or our platforms.
