Tell us the problem you want to solve
Thank you. We have your message and will reply within one business day.
Machine learning that catches what rules miss
We build fraud, credit, and document systems for banks, insurers, lenders, and exchanges, trained on your data and run in your environment.
What we build
Fraud and AML detection
Models that catch fraud, money laundering, and suspicious transactions that rules miss, and keeps catching them as behaviour changes.
Credit and risk
Scoring and approval models built on transaction and credit bureau data, with the explanations your credit committee and regulator need.
Document and knowledge systems
Extraction from statements, claims, and financial reports into a checked data layer that records the source of each fact, so answers can be traced and audited.
Agentic workflows with people in the loop
Agents that prepare the work (a case file, a reconciliation, a draft report) and hand the decision to a person, with language-model usage and its cost visible and capped.
Sovereign and small-model AI
Fine-tuned small models that match large general-purpose models on your specific task, running on your own infrastructure, with your data kept in South Africa.
AI strategy
Senior practitioners who help your leadership choose use cases, set realistic cost and accuracy targets, and build the internal capability to run them.
Built for regulated environments
Explainable by design
Each score comes with the reasons behind it. Where a decision must follow fixed, auditable rules, we derive those rules from the model.
Your data stays yours
Systems run in your cloud or on your hardware. You own the code, the models, and the documentation. You pay no per-seat or per-query licence fees.
People stay in control
Investigators, analysts, and credit officers decide; the system ranks cases, explains its scores, and drafts case files and reports for them to check.
Local and senior
We are a South African team in your time zone, led by people who have built production machine learning since 2014.
AI that costs less to run as it is used more
Renting a general chat model looks cheap at the start and grows more expensive with every question your staff ask. We engineer the data the model reads, pick the most efficient and effective model for the accuracy target, and put it on infrastructure you control, so the cost per answer falls as use grows.
| Rented model, billed per use | Engineered and self-hosted | |
|---|---|---|
| Set-up cost | Low | Higher |
| Cost as users grow | Rises with every user | Flat until a capacity upgrade |
| Cost as questions get harder | Rises steeply | Flat until a capacity upgrade |
| Changing model | Often means starting over | Swap the model, keep the data layer |
Case studySequence learning for fraud detection on a crypto exchange
OVEX's rules caught three quarters of fraud and raised alarms that were mostly false. We replaced them with models that read the order and timing of each account's activity.
We trained two models on OVEX's history, one on how fraudsters behave and one on how honest customers behave, and the system flags an account when it looks more like the first. Now in production, the system includes a dashboard that shows investigators when an account's behaviour changed.
Read the case study →“We partnered with DataProphet for their deep machine-learning expertise and they've built a system tailored to our platform, giving us a far more sophisticated way to identify fraud as it evolves.”
Selected work in financial services
How an engagement starts
Readiness assessment
We review your data and goal and deliver a readiness report, a workshop with your team, and a delivery plan.
Build and validate
We build a working model and measure it against your own KPIs. You then decide whether to go on.
Production and improvement
We deploy, integrate, and keep the models current, and train your team to run them.