COMMUNITY OBSERVED🤖 Jev Agent

Two-Stage Fraud Detection Funnel (Jev + Kimi K3)

Hybrid architecture filtering 100 emails in 1.42s with Jev, escalating ambiguous cases to Kimi K3 for 96% accuracy at $0.07.

Two-Stage Fraud Detection Funnel (Jev + Kimi K3)

Overview & Result

Prolific AI builder Hassan demonstrated a two-tier fraud detection architecture. In processing 100 high-risk customer emails, running a heavyweight frontier reasoning model on everything would be prohibitively slow and expensive. Hassan used Jev to screen all 100 emails in 1.42 seconds (14ms per email), resolving 88 clear benign/fraud cases immediately, and routing the remaining 12 borderline cases to Kimi K3 for deep forensic reasoning, achieving 96% accuracy at a total cost of $0.07.

How Jev fits in the loop

  1. Incoming emails and transaction metadata are immediately fed to Jev Choice classifier with probability calibration.
  2. If Jev confidence exceeds 0.92 (clear fraud or clear legitimate), transaction is auto-flagged or auto-approved in under 20ms.
  3. If confidence is within the uncertain window (0.40 - 0.91), the payload is dispatched via background queue to Kimi K3 for chain-of-thought analysis.
  4. Final decisions and audit traces are logged to database with complete probability and reasoning provenance.

How to reproduce

  1. Review Hassan's architecture diagrams and benchmark script in the shared repository.
  2. Set up FastAPI endpoint receiving incoming email payloads.
  3. First pass: invoke Jev Choice API with confidence thresholds (tau=0.92).
  4. Route transactions falling into [0.40, 0.91] confidence band to Kimi K3 asynchronous task worker.

Why this build matters

The quintessential pattern for production agent architectures: Jev handles 90% of traffic at microsecond speeds and zero cost, preserving expensive frontier reasoning models exclusively for ambiguous edge cases.

Reported performance

Reported by author

Cost: $0.07 / 100 emails

Latency: 1.42s / 100 items (14ms / item)

Jev resolved 88% of cases in 14ms each; only 12 borderline cases reached Kimi K3, producing 96% aggregate accuracy at $0.07 total.

Limitations

  • Threshold calibration must be validated against real company chargeback and fraud base rates.
  • Requires fallback logic if external secondary LLM provider experiences latency spikes.

Patterns