Real-time fraud & risk engine · Lagos, Nigeria
ISAT is a real-time fraud and risk engine for Nigerian digital payments — it scores a transfer the moment it is initiated and shows the reasons behind the score, so banks and fintechs can allow, step up or hold it before the money moves.
Today The risk engine scores a transfer in the browser from transparent rules, and we are training the learned model on cloud NVIDIA H100 GPUs.
Next Real-time scoring at payment latency.
On NVIDIA We are training ISAT's risk model on cloud NVIDIA H100 GPUs: NVIDIA RAPIDS and CUDA run feature engineering across millions of transfers, and a PyTorch graph model learns how money moves between accounts. Next, we will score each payment in real time with NVIDIA Triton and NVIDIA Morpheus.
Why now Nigerian banks and fintechs lost ₦25.85 billion to payment fraud across 67,518 incidents in 2025 (NIBSS).
Works today · Rules-based demo
Describe a transfer and the engine scores it the way the production rules do: a 0–100 risk, the reasons ranked by weight, and the action a bank would take. Everything runs in your browser; nothing is sent to us. This is the rules engine as it stands today — transparent rules you can read — not a trained model.
Risk score · 0–100
Hold
Pause the transfer for review before any money moves.
Reason codes · ranked by weight
Sum of weights 79 → capped at 100 → score 79
Rules-based demo. Runs in your browser; no request leaves this page. Not a trained model and not a decision on any real payment.
Product
Most teams score risk with static rules that fraudsters learn to route around. ISAT gives banks, fintechs, microfinance banks and payment processors a real-time, explainable score — every number comes with its reason codes — built so the learned model now in training can take over from the rules without changing the interface.
Every transfer gets a 0–100 risk and the reasons behind it, ranked by weight: the amount against the sender's pattern, how new the recipient is, velocity, device and hour. An analyst can read the decision, and a customer can be told why.
79 · Hold · R2 recipient under a week old · D1 new device · H2 late-night hour
Static rules are written once and fraudsters learn their edges. The learned model is trained on money-movement patterns instead, so it adapts as new fraud shapes appear — and keeps returning reason codes, not a black box. Training is under way on cloud NVIDIA H100 GPUs; the model is not yet finished or deployed.
In training · PyTorch money-movement graph model · RAPIDS + CUDA features · cloud NVIDIA H100 GPUs
The score has to arrive inside the payment flow, not after it. The target is a decision in under 50 milliseconds, so the transfer is allowed, stepped up or held before the money leaves.
Target · <50 ms · NVIDIA Triton
A static web app: a transfer goes in, a score, ranked reason codes and an action come out. Try it above — it is the same rule table the hero console uses.
What is in training or in development, and not yet part of the demo.
How it works
Four steps, designed to sit inside a bank's or fintech's payment flow. Today the scoring step is a rule table; the learned model now in training will sit behind the same interface, so the reason codes and the three actions stay the same.
The payment request as it arrives, before any money moves.
The transfer is placed against what is normal for this sender.
Each triggered rule adds its weight; the list is returned ranked, largest first.
One of three actions, returned inside the latency budget.
Why now
Industry coordination cut the value lost to payment fraud in 2025 by about half. The number of incidents did not fall with it, transfers remain among the largest loss channels, and the regulator is warning of new fraud patterns as digital payments grow. The next gain comes from smarter, real-time, adaptive scoring — not from more static rules.
₦25.85bn
lost to payment fraud by Nigerian banks and fintechs in 2025.
Source: NIBSS, 2026
67,518
fraud incidents reported in 2025. The attempts did not fall with the losses.
Source: NIBSS, 2026
−51%
value lost versus 2024 (₦52.26 billion), after a year of industry coordination.
Source: NIBSS, 2026
Nigerian banks and fintechs lost ₦25.85 billion to payment fraud across 67,518 incidents in 2025 (NIBSS).
Technology
The demo needs no GPU; it is a static web app. The learned model does: feature engineering over millions of transfers with graph features is GPU work, and that training is running now on cloud NVIDIA H100 GPUs. Serving the model inside a payment's latency budget is the next step.
Static web app, no GPU. Transparent rule table, reason codes, three actions.
Feature engineering and a money-movement graph model, in training on cloud NVIDIA H100 GPUs.
Score and reason codes returned in under 50 ms, inside the payment flow.
Real-time pipeline and graph-based detection over the stream of transfers.
Company
Intelligent Systems & AI Technologies Limited (ISAT) — Lagos, Nigeria.
Incorporated in Nigeria, 2026. Today: the risk engine scores a transfer in the browser from transparent rules, and we are training the learned model on cloud NVIDIA H100 GPUs. Next: real-time scoring at payment latency.
Stage: prototype, pre-revenue. The engine is built for banks, fintechs, microfinance banks and payment processors in Nigeria.
AbdulAzeez Babatunde Adeyinka
Founder
Founded and built by AbdulAzeez Babatunde Adeyinka, Founder.
ISAT has one person on it: the founder designs the engine, writes the code and runs the company.