Verify the work yourself. Don't take our word for it.
Every service Afrispan offers is grounded in a real, working technical capability, publicly demonstrable on GitHub, not a service description written ahead of the capability existing.
Five principles this work has proven, not assumed
Discovered across nine executed phases of real, working evaluation and orchestration, not derived from theory.
An AI system cannot audit itself
The most common form of AI quality assurance available today is the vendor's own claim that its system works correctly. This is structurally unreliable, because an AI system judging its own output shares the same blind spots as the system being evaluated.
Surface pattern-matching fails at the edges
Systems that appear to work reliably on typical cases can fail precisely where the judgment matters most, at the edge cases a governance process is meant to catch.
Monitoring is not governing
Watching a system and having the authority and mechanism to stop or correct it are two different capabilities. A dashboard that only observes is not a governance control.
Authority is not judgment
A mechanism that can halt a system is not the same as one that can decide correctly what to do next. Real oversight requires a named, accountable person exercising judgment, not just a switch.
Judge and human disagreement is diagnostic, not noise
When a cross-model judge and a human reviewer disagree, that disagreement is signal worth investigating, not error to be averaged away. Afrispan's own testing measured a 0.15 point quality inflation and a 0.23 point gap in error detection between a same-family judge and an independent, cross-model evaluator.
Three connected projects, full commit history intact
Published openly at github.com/Afrispan-AI/ai-governance-suite, migrated with full commit history from the founder's original personal repository.
| Project | Scope | Status |
|---|---|---|
| Project 1 | Nine phases proving core governance principles, including a real, working kill-switch with immutable, hash-chained audit logging, and the FRIA, Conformity Report, and Cross-Border Framework generators | Proven Complete, executed against real API calls |
| Project 2 | Thirteen-notebook production evaluation suite, cross-model judging, red-teaming mapped to OWASP and MITRE ATLAS, a verified CI/CD pipeline on GitHub Actions | Building Architecture complete and pytest-verified; most notebooks run in simulated mode pending funded API billing |
| Project 3 | Live governance orchestration in n8n: three-queue PR routing with non-blocking human review, a kill-switch converging drift alarms and red-team findings into one shared gate, pre-committed evaluation verdicts, and a real-time dashboard | Proven Built, tested, and verified on real local hardware, the first phase of this portfolio to reach that status |
Project 3's pre-committed verdict design closes a real gap named by an external reviewer, Federico Blanco Sanchez-Llanos, before that gap could reach a client engagement.
What this is actually built on
Evaluation and observability
RAGAS, DeepEval, and Langfuse power the test harness and drift monitoring layer, real evaluation scoring, regression alarms watching for degradation over time, and full observability tracing behind every judgment made.
Red-teaming
Promptfoo drives adversarial payload testing and red-teaming, mapped to OWASP and MITRE ATLAS.
Governance orchestration
n8n handles governance orchestration and human-oversight enforcement.
CI/CD verification
GitHub Actions verifies the CI/CD pipeline.
Engineering delivery
Claude Code delivers the engineering itself, with the same discipline of small, verified increments and honest status tagging used throughout Afrispan's own technical foundation.
Original repository
The founder's original personal repository at github.com/steveonyeke/python-ai-governance carries a note pointing to the current organisation repository and remains archived, not deleted.
Go look at the actual code
A technical stakeholder should be able to verify this work directly, not take a claim on faith.