Tell us the problem you want to solve
Thank you. We have your message and will reply within one business day.
The right AI
for the job
We build and run custom machine learning for financial services, manufacturing, and other businesses. We choose the most efficient and effective model for your accuracy target and run it in your environment. You own the code.
Most engagements start with a fixed-fee readiness assessment. It takes four to eight weeks and ends with a written recommendation on what machine learning can do for your goal and what it would cost to build and run.
of fraud caught at OVEX, a South African crypto exchange, against 75% for its rules, with no language model.
in production for the fraud scoring and investigation system we built for a multinational insurer.
measured annual uptime across 380 sites on DataProphet Connect, our managed industrial data platform.
Plant optimisation
Prescribe finds the process settings that cut scrap and defects, and runs on DataProphet Connect, the industrial data platform at hundreds of plants in over 30 countries.
Learn more →AI engineering
Custom AI systems for finance and back office, fraud and risk, documents and knowledge, engineering design, demand forecasting, and AI strategy, for businesses in any industry.
Learn more →Working on something else? We have also built machine learning systems for retail, telematics, health, legal, and gaming. Let us work on your challenge.
We start with the problem, not the technology
We start from the number you need to move: fraud losses, scrap rate, hours spent on manual review, cost per decision, etc. Then we work back to the simplest system that achieves your goal. If the data is not ready, we tell you what needs to change before you spend on a build.
A frontier language model suits some problems. For many others a smaller or classical model is as accurate or better and costs less to run. Because we build machine learning from first principles, we choose from the whole field: statistical models, classical ML, deep learning, fine-tuned small language models, and frontier models, alone or combined. You get the accuracy you need at a running cost you can predict.
Two hidden Markov models and a risk score, no language model. Catches 98% of fraud.
Read the case study →For Shoprite's micro-lending business, a model declined applications that were unlikely to succeed before the paid credit bureau call was made, saving about 20% of bureau costs.
We benchmark document AI, LLM extraction, and hybrid methods on your documents and choose on accuracy and cost per page.
Recent work
Financial services · 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.
Read the case study →
Manufacturing · Case study Near zero-defect quality at a tier-1 engine-block manufacturer Internal scrap halved in month one, about $100k saved per month. Read the case study →
Manufacturing · Case study Scrap reduction at a global light-alloy wheel manufacturer Casting scrap cut by 29% on one wheel, projected 10x two-year ROI. Read the case study →In our clients' words
“In DataProphet, we have found a partner who shares our pragmatic attitude and our passion for helping foundries work ever more productively and resource-efficiently.”
“This is why I rate you and your team: no fluff, just great theory applied.”
“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.”
Tell us the problem you want to solve
Most engagements start with a short readiness assessment: we look at your data and your goal, and tell you plainly what machine learning can do for it and what it would cost.