Attesta (sold as GRCAssure in Nigeria)
Evidence-native enterprise GRC platform for banks and regulated entities in Africa.
Context
Regulators expect banks to prove that controls work, not just that a policy exists. Most teams still run this on spreadsheets and email. A regulated financial institution evaluating a proof of concept also needed assurance that its data would not be exposed through AI features.
My role
Founder and lead engineer: product, architecture, data model, AI integration, security review and cloud deployment.
What I built
Risk register with 5x5 scoring and four treatment routes (reduce, accept, transfer, avoid). Assessment engine for applications and vendors. Compliance mapping and posture across a 10-framework library including ISO 27001 Annex A and ISO/IEC 42001. AI policy drafting and scope-aware policy review. Remediation with independent verification. DPIA. AML/CFT/CPF compliance programme for CBN-regulated entities. Business continuity (BCP and BIA). AI governance register. Internal audit.
Engineering decisions
- 01
Private AI gateway
All nine AI features call one edge function that forwards the caller's auth to AWS API Gateway, then a Lambda authenticated by a shared secret held in Secrets Manager, then Claude on Amazon Bedrock. A metadata-only invocation log records who called which capability without storing prompt content. Per-organisation, per-feature metering caps cost, and a capped user gets a clear message while every feature still works by hand.
- 02
Deterministic scoring, AI on top
Branching questionnaires score 0 to 100 internally, hidden branches are excluded and the remaining weights renormalised, and results display on a 1 to 5 scale. The AI writes the narrative and recommendations but can never change the score, so every rating is reproducible for an examiner.
- 03
Status is earned, not typed
Control implementation status is derived by database triggers from sign-off, a named owner and operator, and evidence. Uploading evidence submits for verification, and only a reviewer independent of the owner and operator can close it. Residual risk moves only when remediation is verified closed. Accepted risk is shown as carried, never as mitigated.
- 04
Multi-tenant security
Every table has row-level security scoped to the organisation, and the audit log and residual snapshots are append-only, enforced by triggers. I ran structured audits of access policies and error-message leakage, and closed a cross-organisation read path on policy approvals.
- 05
Grounded generation
The policy generator and the policy reviewer share one grounding: only the selected framework's requirements and mapped controls, bounded by the policy's scope category, so an access-control policy does not drift into cryptography or backup. Policy statements are written as commitments, never as claims of current practice.
Architecture
Stack
Status and links
Live. Enterprise proof of concept under way. Target state: in-country hosting with a dedicated per-client AI deployment.