Composite patternGoogle CloudReal-time ML

High-volume risk decisions with an explanation attached.

This lending pattern shows how lower latency and better approval quality can coexist with feature consistency and an inspectable decision path. Benchmarks below are illustrative targets pending client-approved evidence.

18MScores per day
82msP95 latency
23%Risk capture lift
6 minModel rollback
01 · Constraint

Online speed was useless if features meant something different.

Batch training features were rebuilt in application code for serving, producing subtle skew. New borrower segments also drifted faster than the quarterly review cycle could detect.

Feature skew

Training SQL and online transformations had separate owners and release paths.

Opaque decisions

Operations received a score without the evidence needed for exception review.

Slow rollback

Endpoint changes and feature changes could not be reversed as one unit.

02 · Decision path

One feature definition served training and inference.

IngestStreaming signals

Repayment, device and application events

DefineFeature views

Point-in-time correct offline and online data

TrainVertex pipeline

Reproducible training and bias checks

ServeVertex endpoint

Low-latency score and reason codes

WatchDrift + outcomes

Segment, feature and approval monitoring

Failure path: feature freshness or fairness breaches routed traffic to the previous signed endpoint and opened a review with the affected borrower segment attached.
03 · Decisions

Explainability entered the API contract.

D-01Point-in-time joins

Training examples could only use data available at the historical decision moment.

D-02Reason codes at inference

Each score returned stable factors suitable for operations and adverse-action review.

D-03Segment-aware monitoring

Aggregate quality could not hide drift or unfair impact in a smaller cohort.

04 · Outcome

Faster decisions with a safer model lifecycle.

MeasureBeforeAfter
Prediction latency310 ms p9582 ms p95
Daily scoring capacity6.5M18M
Training-serving skewRecurring incidentsShared feature views
RollbackHours6 minutes
Vertex AIVertex AI PipelinesBigQueryDataflowPub/SubFeastCloud Monitoring
Risk principle: the production unit was not just a model. It was model + feature definitions + explanation schema + segment thresholds.

Build real-time ML without creating a blind spot.

We can map your feature contract, latency budget, explainability needs and rollback boundary.

Review my GCP ML path →