Real-Time Risk Scoring Engine
A lending platform running batch overnight risk checks was losing deals to faster competitors. Every hour of delay in a loan decision was a customer who had already signed with someone else.
A consumer lending platform processing millions of loan applications across digital channels.
Batch-based risk processing created hour-long delays in loan decisions, causing drop-off and increasing default rates from under-reviewed applications.
Quick Summary
We replaced a lending platform's overnight batch risk checks with a real-time ML scoring layer built on Python and MindsDB, fed by a streaming AWS Kinesis pipeline. Loan decisions now take under 2 seconds instead of up to 8 hours, default rates fell 40%, and the platform handles 10× the application volume without added infrastructure.
Challenges Identified
Overnight batch runs meant loan decisions took up to 8 hours — an eternity in a market where competitors were approving in minutes.
Risk models were static — trained quarterly and unable to adapt to shifting borrower behaviour between refreshes.
Manual review queues bottlenecked at underwriters — edge cases and flagged applications piled up, slowing throughput across the board.
No unified data pipeline — credit bureau data, transaction history, and behavioural signals sat in separate systems with no real-time aggregation.
Sakha Solution
ML-driven real-time scoring layer built on Python and MindsDB — evaluates each application against live model predictions the moment it arrives.
Streaming data pipeline via AWS Kinesis — aggregates credit signals, transaction history, and behavioural features in sub-second latency.
Adaptive model retraining cadence — automated retraining triggers when model drift is detected, keeping predictions current without manual intervention.
Automated decisioning rules engine — routes applications to approve, decline, or human review based on configurable risk bands, eliminating most manual queues.
Technical Highlights
Impact
Measurable outcomes delivered to our client.
Loan decision latency collapsed from hours to under 2 seconds — a step-change in conversion on digital channels.
Default rate fell 40% — real-time signals caught risk patterns that batch models consistently missed.
Underwriter review load cut by 65% — automation handled the volume, humans focused on genuinely ambiguous cases.
Platform handles 10× the application volume without adding infrastructure — the scoring layer scales horizontally on demand.
"We went from overnight batch runs to live decisions in seconds. The difference in conversion was immediate."