The Playbook to Stop Polymorphic AI Fraud
Author detection logic in plain language
Create and edit rules, features, and lists in natural language — turning fraud intent into validated, production-ready detection logic in minutes, without engineering bottlenecks.
Write it in words. Ship it as logic.
Vera turns plain-language intent into validated rules, features, and lists — quick-tested before anything ships.
Write rules the way you think
Create, edit, explain, and generate variants of rules from plain language, composed with feature-operator expressions.
Every feature type, checked before review
Aggregation, non-aggregation, Java code, label, synthetic joiner, list, and nested features — validated against the Feature Platform APIs before you ever see them.
Manage lists and test before rollout
Create and validate lists inline and quick-test on production data (the latest 500, up to 1,000 events) before anything ships.
From instruction to published entity
Type an instruction, Vera drafts and validates it against the platform APIs, and you quick-test on recent events before anything ships.
From instruction to published entity
Type an instruction (e.g., alert when a metric exceeds a threshold).
Vera drafts and validates it against the platform APIs.
Review it on screen and quick-test on recent events.
Approve.
The entity enters the standard draft-and-publish lifecycle.
Validated objects, not free-text suggestions
Outputs are validated against platform APIs, then held for human review through a four-stage lifecycle, RBAC-scoped to what the user can already touch, and traceable to the originating chat session in the audit log.
Detection logic across rails
Rule creation and updates, feature creation, list creation and editing, and quick testing — for fraud and AML detection logic across rails.
Weeks to minutes
In one deployment, a reversible risk switch built without rewriting 100+ core rules cut false positives for a merged cohort by roughly 80% and shipped in under a day.
Build big changes without the rewrite
During a merger, a credit union used Vera to design and deploy a reversible risk switch without touching its 100+ core rules — then returned members to standard treatment once the risk passed. “Having Vera is like having another one of me on the team. You tell Vera what you want to implement, it builds it, and you validate it.” — Fraud Specialist
Real, governed platform entities
Vera validates against the Feature Platform APIs and creates real, governed entities using your own data and configurations — resolving list dependencies and quick-testing against production events, which a generic LLM cannot do.
Governed, private, and auditable
Like every Vera agent, rule and feature creation runs on Amazon Bedrock statelessly — prompts and data are never stored by AWS or used to train foundation models, and are processed in full per-customer isolation.
- Stateless processing on Amazon Bedrock — nothing stored by AWS
- Four-stage lifecycle: Pending Approval → AI Generated → Accepted → Published
- RBAC-scoped to what the user can already touch
- Traceable to the originating chat session in the audit log
- Clear, single-objective prompts with real field names produce the best results
Common questions
What’s in scope today?
Features, rules, and lists plus quick testing — not backfill, profiles, entities, or scheduled runs.
How accurate is it?
Vera can misinterpret intent, so all outputs are reviewed before applying; clear, single-objective prompts with real field names produce the best results.
Turn intent into detection logic
See how Vera authors validated rules and features from plain language.