Build and own a Promotion Abuse Specialist Model.
Is this promotional redemption, coupon use, or referral legitimate, or does it represent abuse that should be blocked, clawed back, or investigated?
Overview
Reduce promotional spend lost to coupon stacking, bot redemption, and referral fraud.
Who owns this decision
VP Trust & Safety, Head of Growth Marketing
Why owned AI may fit
Company-specific promotion rules and eligibility terms are not represented in any general model's training data.
What this Specialist Model actually does
Decision: Is this promotional redemption, coupon use, or referral legitimate, or does it represent abuse that should be blocked, clawed back, or investigated?
This is a larger system with 1 specialist model candidate and 5 supporting or external components.
Who uses this Specialist Model
Accountable owners: Promotion and growth, Trust and safety, Fraud and risk, Payments and finance, Privacy and legal, Customer support and appeals.
- Promotion owner
- Growth owner
- Trust and safety analyst
- Fraud analyst
- Risk team
- Customer support
- Payments
- Finance
Where the specialist model sits
This is a larger system with 1 specialist model candidate and 5 supporting or external components.
Fine-tuned classifier
Specialist Model
Fine-tuned classifier for abuse-type categorization
Graph-based linkage analysis across accounts, devices, and payment methods
Supporting component
How Oumi connects to your stack →
Rules engine
Supporting component
Rules engine for promotion-specific eligibility checks
How Oumi connects to your stack →
Real-time scoring pipeline at redemption time
Supporting component
How Oumi connects to your stack →
Analyst review queue and case-management interface
Supporting component
How Oumi connects to your stack →
Continuous retraining process using confirmed abuse and chargeback outcomes
Supporting component
How Oumi connects to your stack →
Decision and authority boundary
Promotion identity account device email phone network bot payment order shipping fulfillment return behavior graph risk eligibility calculation loss action appeal remedy and human systems establish factual financial and action truth. Oumi leads term and case extraction evidence linking contradictions gaps typology and alternative hypotheses case synthesis explanations communications confidence abstention and routing.
Abstain or escalate when
- Merchant missing
- Promotion version missing
- Campaign missing
- Jurisdiction missing
- Channel missing
- Customer missing
- Account missing
- Redemption missing
How to evaluate and improve it
Feedback signals
- Oumi hypothesis
- Evidence links
- Alternatives
- Deterministic result
- Analyst finding
- Correction
- Approval
Model candidates to evaluate
- ModernBERT large (Apache 2.0)
- Qwen3.5-4B (Apache 2.0)
- NeoBERT (MIT)
These are candidates, not performance claims. Run them against your evaluation set.


