mainza-ai/afreval

Open-source AI evaluation for Africa. A pre-deployment compliance harness for AI in Africa — evidence-linked Context Scores and per-language profiles, open source, built on WAXAL · AfroBench · afri-fertility. # Milimo AfrEval **Open-source AI evaluation for Africa — a pre-deployment compliance harness that gates agents with an evidence-linked Context Score and per-language Context Profile, built on WAXAL · AfroBench · afri-fertility.** **Author:** Mainza Kangombe AfrEval routes every AI agent through a zero-trust sandboxed certification pipeline before it touches a production database, producing an **evidence-linked Context Score** plus a per-language/per-script **Context Profile** that measures how linguistically accurate, culturally safe, economically viable, and functionally secure the model is — inside a zero-trust sandbox. > **What it is / what it isn't:** AfrEval produces an auditable, evidence-linked evaluation signal — it is a *credible signal and compliance harness*, not a government certificate. Authority comes from ecosystem adoption; the score is the measurement, and the methodology behind it (which probes, which judges, which corpora) is published in every cert. Built on the generalized **Karpathy Loop** (frozen harness → mutable artifact → instruction file → metric → budget), with every subsystem — tokenizer search, Context Score calibration, BiasScope probing, edge ASR, execution-boundary hardening, compliance — running as an autonomous loop over three frozen data substrates: WAXAL, AfroBench, and afri-fertility. ## Assets & Presentations Visual assets and presentations detailing the AfrEval architecture and Secure AI Gateway: - **Video Explainer**: Building the Milimo AfrEval Tollbooth - A Zero Trust Architecture for Context Aware AI - **Benchmarking Infographic**: AfrEval_African_AI_Benchmarking_Infographic.png — High-resolution diagram illustrating African AI benchmarking, zero-trust sandbox execution, and the Context Score pipeline. - **Secure AI Gateway Slide Deck**: - PDF Deck — Complete 15-slide presentation on the AfrEval Secure AI Gateway architecture. - PowerPoint Deck (.pptx) — Editable slide deck. - Exported Slide Images — Individual …