Fraud Screening With Sprappy Filter and SPRAPP Panel
A fintech pattern that uses Sprappy Filter as a pre-LLM gate and SPRAPP Panel to reason about borderline fraud cases.
Fraud Moves Fast
Fraud screening is adversarial by nature — the input is actively trying to fool you. That makes the pre-LLM boundary the most important part of the pipeline. The SPRAPP suite puts Sprappy Filter at that boundary. SPRAPP is a tool, not a compliance officer, and a human owns every decision.
The Pre-LLM Gate
Sprappy Filter scores inbound content across 25 threat categories before any reasoning model sees it. In a fraud context, that means catching injection attempts, manipulation, and adversarial payloads aimed at steering a downstream model. Filtering first means the reasoning layer only ever reasons over content that has cleared a threat check.
Reasoning About Borderline Cases With Panel
Clear-cut fraud and clearly-legitimate transactions do not need a panel. The borderline cases do. SPRAPP Panel reviews a flagged case and reports where models converge.
- Agreement on "likely fraudulent" supports holding the transaction for review.
- Disagreement signals genuine ambiguity that a human must resolve.
A panel that disagrees is protecting you from false confidence — exactly what you want when money is moving.
Compact Case Records With smoltext
Fraud systems store volumes of short strings — transaction IDs, rule codes, status tags. smoltext compresses these short strings efficiently where gzip wastes space on small payloads, keeping the case history compact for the long retention fraud investigations require.
Offline Investigation Notes
Investigators sometimes work in restricted environments. TinyLM's on-device models let them draft case notes with no network, keeping sensitive details local.
Clear Division of Labor
Filter guards the boundary, Panel reasons about the hard cases, smoltext keeps records lean, and TinyLM supports offline work. Humans decide.
Implementation
Stand up Sprappy Filter as the pre-LLM gate first; it is the single highest-value piece in an adversarial setting. Add Panel review for borderline cases once the gate is solid.