Embedded Finance AI Stack for an African BaaS Provider — Powering Lending for 240 Partner Brands
OnboardX + Fraud Guard + Credit underwriting accelerators packaged into a single API the embedded-finance provider's 240 partners consume.
The challenge
The client — an African embedded-finance provider that supplies banking-as-a-service rails to consumer brands across e-commerce, ride-sharing, gig-economy and retail across multiple African markets — was scaling faster than its in-house AI capability could keep up with. The client's partner brands (240 active, growing) integrated with the BaaS API to offer co-branded financial products (savings wallets, instant credit, deferred-payment instruments, prepaid cards) directly inside the partner's customer experience.
Each financial product needed its own KYC flow, fraud screening, credit underwriting, transaction-monitoring and regulatory reporting. The client's small data-science team had built ad-hoc models for each product, with the engineering burden of model maintenance and the regulatory burden of model documentation scaling linearly with the partner-and-product count. The team was at breaking point.
The constraints were structural. The BaaS API had to handle the latency profile of being embedded in consumer-app checkout flows — onboarding had to complete in under sixty seconds in-app, credit decisions had to be returned in sub-three-second latency, fraud screening had to happen in real-time without blocking the partner's transaction flow. The product set spanned multiple regulatory regimes (different country regulators in each African market) with different KYC, AML and credit-disclosure requirements.
The approach
MindMap deployed an AI stack composed of OnboardX (Ox) for the per-product KYC flows, Fraud Guard (Fg) for the real-time fraud-and-screening layer, a domain-specific credit-underwriting accelerator built on Coding Assistant patterns, Compliance Engine (Ce) for the per-jurisdiction regulatory rules, and LLM Gateway (Lg) for the model-management and per-partner cost-attribution layer.
Phase one was the multi-tenant architecture. The client's partner brands required strong isolation — Partner A's customer data must not be visible to Partner B, models trained on Partner A's data should not benefit Partner B unless contractually permitted. The platform was built as a multi-tenant stack with per-partner data isolation enforced at the storage, compute and model-serving layers.
Phase two was the per-product configuration layer. Rather than building separate models per product, we built a configuration-driven product factory: a new partner-and-product combination is onboarded by configuring the KYC document set, the fraud-screening profile, the credit-underwriting model parameters and the regulatory-rules-set, with the underlying accelerators serving the configured product without bespoke code.
Phase three was the per-jurisdiction regulatory layer. Compliance Engine encodes the KYC, AML, credit-disclosure and consumer-protection requirements for each of the African markets the client operates in. Each partner-and-product combination's runtime behaviour is gated by the relevant jurisdiction's rule set — meaning a single partner brand operating in three countries gets the right KYC and disclosure behaviour automatically per market.
Phase four was the operational telemetry layer. The client's partner brands consume detailed telemetry on their financial product's performance (originations, approval rate, default rate, fraud rate, customer-experience metrics) through a partner-portal dashboard powered by the platform's analytics layer.
The pre-built building blocks
Rather than commission a ground-up build, the engagement leaned on MindMap's pre-built accelerator library — production-tested components that compress what would otherwise be a six-to-nine-month build into weeks.
OnboardX
Per-product configurable KYC flows
Fraud Guard
Real-time fraud-and-screening engine with per-partner tuning
Compliance Engine
Per-jurisdiction regulatory rules library
LLM Gateway
Multi-tenant LLM access with per-partner cost attribution
DocuMage
Document-intelligence underneath the KYC flows
The architecture
The platform runs on the client's AWS multi-region footprint, with regional deployments in each of the major African markets the client operates in to satisfy data-residency and regulatory-licensing requirements. Cross-regional data flows are constrained to the minimum the platform's operations require.
OnboardX handles the per-product KYC flows. The KYC document set per product is configurable — different products have different KYC depths, from light-touch wallet onboarding (national ID only) to full credit onboarding (national ID, proof of address, income evidence, employer verification). DocuMage performs the document-intelligence layer, FaceMatch performs the liveness check, and the national-ID database integrations (where available per country) provide the third-party verification step.
Fraud Guard runs as a streaming fraud-and-screening engine on every transaction. The fraud model is a per-partner tunable XGBoost-and-graph ensemble — each partner's fraud profile differs (a ride-sharing partner sees different fraud patterns than a retail partner), and the per-partner tuning has materially outperformed a single shared model. The real-time decision latency stays below 80ms p99.
The credit underwriting layer is a per-product gradient-boosted-tree model with a shared feature store. Features include the partner's first-party customer data (with explicit partner consent), the client's transaction-history data across its full BaaS footprint, and per-country credit-bureau data where available. The model is served at sub-300ms p99.
The LLM Gateway is the access layer for the platform's LLM components — the document-extraction models, the case-narrative generators, the customer-correspondence drafters. The gateway handles per-partner cost attribution (each partner sees their cost-of-AI-services in the partner portal), per-partner rate limiting, and the safety-and-compliance guardrails on all LLM outputs.
The numbers behind the story
240 partner brands are live on the platform, with approximately 1.1 million monthly originations across the partner portfolio (a combination of new account openings, credit applications, instalment-product enrolments). Onboarding completes in a median 58 seconds in-app, credit decisions return in a median 1.8 seconds, and fraud screening adds median 22ms to the partner's transaction-flow latency.
Straight-through underwriting rate is 82% across the portfolio — meaning the platform makes the credit decision end-to-end without human intervention on 82% of credit applications. The remaining 18% go through a hybrid model with light human review on the borderline cases.
The configuration-driven product factory has enabled rapid partner onboarding. The client's average time-to-launch for a new partner brand has dropped from approximately 14 weeks (under the previous per-product-bespoke-build model) to approximately 4 weeks (under the new configuration-driven model). The client's partner-acquisition pipeline has accelerated correspondingly.
Per-jurisdiction regulatory outcomes are consistent. The client's regulatory engagements across the African markets it operates in have not raised material findings against the platform's AI components since go-live, with the per-jurisdiction Compliance Engine rules ensuring per-market regulatory adherence by construction.
The client's in-house data-science team has been transformed. The team has shifted from building per-product models to operating the platform — building configuration templates, monitoring model performance across the portfolio, and partnering with new partner brands on their product design. The team's headcount has grown modestly but per-team-member partner coverage has grown substantially.
“Our partner growth was being throttled by the engineering and regulatory burden of standing up a bespoke AI stack per partner. MindMap's configurable platform broke that constraint. We are now onboarding partners in four weeks instead of fourteen, and our regulatory engagements across the African markets we operate in have been consistently positive. The platform has changed the shape of our growth curve.”— Chief Technology Officer· African Embedded Finance Platform
Why MindMap was chosen
The client had previously been building each partner brand's AI stack as a bespoke engineering effort, with the engineering and regulatory burden scaling linearly with the partner count. The CTO had concluded that the linear-scaling pattern was the binding constraint on the client's growth.
MindMap's accelerator-composition approach — bringing OnboardX, Fraud Guard, Compliance Engine and LLM Gateway together into a configurable platform that scaled with partner count rather than with bespoke engineering effort — was the structural differentiator. We could demonstrate the multi-tenant pattern at another embedded-finance context (a different industry, but the same architectural pattern).
Our multi-jurisdiction African regulatory expertise — three former African banking regulators on the delivery team plus the cumulative experience of MindMap's African banking deployments — was the regulatory differentiator. The client's CCO felt that the team understood the per-country regulatory reality of African embedded finance, not just the technology.
Related deployments
Sovereign WhatsApp Banking
ChatNext-powered WhatsApp bot deployed inside the bank's air-gapped data centre, handling balance, transfers, statements and loan applications in English and Swahili.
Cheque OCR at 99.1% Accuracy
DocuMage replaced a legacy template OCR for cheque clearing, processing 10,000 cheques per day at 99.1% field accuracy with 94% straight-through processing.
UK Challenger Bank KYC
OnboardX rebuilt the KYC pipeline with liveness, sanctions and PEP enrichment in a single STP flow, collapsing onboarding from 5 days to 4 hours.
Want an outcome like this?
Start with a 2-week AI Readiness Sprint. We deliver a prioritised use-case backlog and business case grounded in what's actually buildable with our accelerator library.