Overview

Designed a fraud detection workflow for fintech operations with risk scoring, review queues, and alerting.

Problem

Teams in fintech needed a system that could support real production demands without drifting into fragile workflows, unclear ownership, or hard-to-debug failures. The goal was to ship something that felt practical to operators, maintainable to engineers, and credible to business stakeholders.

Approach

The work was structured around explicit state handling, clearer operational visibility, and architecture choices that would still make sense once load, edge cases, and maintenance pressure showed up in real life.

Why this approach

Designed layered risk scoring for transaction screening.

Architecture

• Designed layered risk scoring for transaction screening.
• Added human-review support with clearer investigation context.
• Tracked explainable fraud signals for compliance-facing teams.

Stack

• Python
• PostgreSQL
• Kafka
• Feature Flags

Trade-offs

• Preferred maintainability and operational clarity over clever abstractions.
• Kept workflows explicit so support and product teams could understand system behavior.
• Chose gradual rollout paths instead of risky all-at-once changes.

Outcomes

• Improved suspicious activity prioritization.
• Reduced manual review overhead on low-signal cases.

Lessons

• Good systems are easier to operate when state and ownership are visible.
• Reliability work becomes easier when product and operations can inspect the same truth.
• The best technical decisions usually reduce both engineering risk and business confusion.