Transaction risk assessment
A probability for every transaction from a supervised model, with the factors behind it.
fraud prevention · explainable ML
TrustSignal is a proposed fraud detection platform for digital businesses. It scores every transaction with machine learning and your own rules, shows exactly why, and holds the risky ones for your investigators — who make the call.
Every assessment comes with its evidence: which factors pushed the risk up, which brought it down, and by how much. This is the prototype’s real model, trained in your browser on synthetic data.
Contribution in log-odds relative to an average transaction. Red raises risk, green lowers it.
A lower threshold catches more fraud and holds more good customers. TrustSignal shows both, measured on 300 held-out labelled transactions — so you set policy knowing the cost.
A probability for every transaction from a supervised model, with the factors behind it.
Sequences such as a new login, a password reset and a payout change within a day, flagged as one alert.
Amount against the account’s usual, velocity, device, location and time — compared with each account’s own normal.
Held volume, labels, agreement between model and investigators, and money at risk.
Review and priority thresholds plus readable business rules you switch on and off.
Group related holds by account, add notes, label outcomes and close with a summary.
One call per transaction returns the score, factors, matched rules and recommended handling.
Risk assessments accompanied by understandable supporting evidence, with human review and correction mechanisms built in.
Wallets, cards and transfers where speed matters and every false decline costs a customer.
Many sellers and buyers, new accounts every day, and payouts that must reach the right people.
Card testing, trial abuse and account takeovers spotted early.
decision-makers Risk officers · fraud prevention managers · payment operations executives
Recurring subscriptions with usage-based transaction analysis fees. Illustrative pricing for a proposed product.
TrustSignal is a proposed fraud detection and transaction risk platform for fintech companies, ecommerce marketplaces and subscription platforms. It must score transactions in real time, explain every score, and keep reviewed decisions as durable records for retraining and audit. Designing on AWS lets it train and host supervised models, process transactions as they happen and expose a low-latency risk API.
Amazon API GatewayAWS Lambda
Transactions arrive through the risk assessment API and are enriched with account and device features.
Amazon SageMaker AI
A supervised risk model scores each transaction; configurable rules run alongside it.
Amazon DynamoDB
Held transactions, investigator labels and corrections are stored as reviewed risk records.
Amazon Bedrock
Case summaries are drafted from the transactions, factors and notes for investigators to edit.
Amazon S3
Labelled training datasets are versioned so models can be retrained and evaluated for false positives.
TrustSignal is a proposed AI fraud prevention and transaction risk assessment platform for fintech companies, ecommerce businesses and digital marketplaces. The planned application will identify unusual activity patterns, generate risk assessments and organise cases for human investigation. We plan to use Amazon SageMaker AI for predictive models, Amazon Bedrock for case summaries, Amazon S3 for datasets and Amazon DynamoDB for reviewed case records. AWS Lambda and Amazon API Gateway will support processing and application integration. Risk assessments will be evaluated for false positives, with human review and appropriate correction processes.
Planned, not yet deployed. This website and the risk desk run in your browser on synthetic data.
Peter founded TrustSignal after watching payment teams choose between black-box fraud scores they couldn’t explain and rules that blocked good customers. Peter leads product, customers and every pilot, with a focus on explainable scores and measured false positives.
TrustSignal is a proposed product in its building stage. A demo walks through scoring, explanations, thresholds, the review queue and cases on synthetic data. The MVP covers synthetic transaction datasets, anomaly classification, configurable review thresholds and investigator-approved case records.
By Peter Olamide, to give fraud teams risk scores they can understand and correct.
Supervised model plus configurable rules, factor-level explanations, review thresholds and investigator corrections.
Synthetic transactions, explained scores, thresholds with precision and recall, review queue, account alerts and cases.
Model training and hosting, processing, risk API, reviewed records and case summaries designed.
Shadow-mode scoring on a customer’s historical, labelled transactions before any live review queue.
No. It recommends handling — approve, or hold for review. Investigators make every decision on held transactions.
Every threshold shows its precision and recall on labelled data, investigator labels are recorded as corrections, and models are re-evaluated before they change.
Transaction and account behaviour — amount against the account’s usual, velocity, device, location and time. The prototype uses seven synthetic features.
Not yet. It is a proposed product; the risk desk is a prototype running in your browser on synthetic data.
Questions about the product, the API and integrations, data handling, or partnerships. Messages go directly to the founder.