Supercharge your predictive AI projects.

Example project workflow

​

The following example workflow – for a fraud detection project – illustrates the importance of Gooder AI's comprehensive functionalities. See also the FAQ "Why should my team use Gooder AI instead of our own DIY approach?"

​

Customize for your project. To define your main metric – savings – you set fraud loss (and fraud-prevention gain) to be a proportion of each transaction's magnitude, and the false-positive cost to be a flat cost between $50 and $150. You also set up the display according to project particulars, using domain-specific language – e.g., the model decides "which transactions to block." [Demos: video 1, video 2]

​

Zoom in. The savings curve shows one of your three models dominates, delivering and estimated $22 million in savings per quarter. You zoom in to the top 5% of cases, since your deployment will probably only block at most a couple percentage of the highest-risk transactions.  This reconfirms the leading model's dominance. [Demo video]

​

Managing uncertainty in business factors. You know that your orgnization may be on the hook for only a fraction of each fraudulent transaction's loss, so you reposition the corresponding slider. You can see that it has a significant impact on the saving curve's shape, but, so long as it is within the range of 70% to 100%, it makes very little difference as far as both the choice of model, as well as the best position for the decision boundary (corresponding with the curve's peak). [Demo video]

​

Comparing competing KPIs. For peak savings, the number of wrongly-blocked, inconvenienced customers shows as 85k per quarter. This is already accounted for in the savings estimate, but there may be intangible or longer-term costs associated with it, such as the reputation of your institution and the adverse effects of disrupting commerce. You can see that the profit curve is relative flat around the peak. This means you could move the decision boundary toward the left – that is, block fewer transactions – with very little decrease to savings. By trying this interactively, you find that, by blocking only 2.3% of transactions instead of 3%, the number wrongly blocked is cut by a whopping 50%, while only decreasing the estimated savings from $22 million to $21.5 million.

[Demo video (same as above)]

​​

Vetting a proposed initiative. Your executive team has been discussing an "apology gift card" treatment every time a legitimate cardholder suffers the inconvenience of a false-positive interruption. In the most optimistic scenario, this might decrease the net cost down from $75 to $50. However, changing that slider accordingly shows that this would only increase the estimated savings from $22 million to $24 million. It also shows that reaching maximal savings would require more blocked transactions: 3.5% rather than 3% (interrupting more commerce). Given the complexity of this proposed initiative, the uncertain results, the lack of a great upside, and the lack of precedent, your organization decides to dispense with the proposal. [Demo video (same as above)]

​

Targeting by expected value. Would fraud detection deliver more value if it targeted for blocking not simply the transactions more likely to be fraudulent, but instead the transactions expected to cost the most due to fraud – that is, those for which the expected value of blocking would be highest? You add this as a derived data column that multiplies fraud risk (model score) by transaction size. Now you set that new column as a "model" to compare with the "raw" models. The results are remarkable: Within the savings curve, you can see new deployment options that are much more compelling, such as attaining an expected savings of $24 million ($2 million higher) by only blocking half as many transactions (1.5% rather than 3%). [Demo video]

​

Driving model-development toward business value. A data scientists delivers a new competitive model based on a method that everyone thought would help, but its technical performance doesn't seem to be much better – it has about the same AUC as your current champion model. However, when viewed within Gooder AI in terms of savings, it turn out to be the new champion – and when used as part of the expected value calculation, it improves that even further!

​

Business stakeholder usage. Your customer wants to explore the Gooderboard herself and get a feel for the tradeoffs offered by various deployment options. You share your current setup with her via Gooder's cloud-based view-sharing capability.

​

Sub-segment views. Your customer follows many of the same interactions above, getting a feel for how deployment options affect the estimated value. She selects cross-sectional views for different geographical regions, finding some where model performance is lower. This spurs an investigation into potentially varying the deployment by region.

​

Gooder AI chatbot. Your customer is new to savings curves and decision boundaries. Although she seems to catch on fairly quickly, she greatly appreciates the ability to ask endless questions of a "virtual data scientist" – without ever fearing she's pestering, overtaxing, or asking “stupid questions.” [Demo video]