Brazil – October 27, 2026 -- Snowflake has made its Online Feature Store generally available, serving machine learning features in under 10 ms for real-time fraud detection. Two financial-sector customers, Neon and VALID Systems, report lower latency and lower fraud losses after moving to the platform.
Neon halves P99 feature latency to 72 ms after vendor discontinuation
Neon, a Brazilian digital financial institution with over 27 million customers, routes every login, card payment, transfer and loan request through Viper, its in-house fraud engine. Its previous feature store vendor, Tecton, was being discontinued, which set a hard deadline.
In three months, Neon's team rebuilt almost 300 features across eight fraud workflows on Snowflake Feature Store. P99 feature latency fell from 140 ms to 72 ms, almost twice as fast. Viper's service availability rose from 99.7% to 99.9%.
Neon cited low latency, fresh streaming and real-time features, identical governance across online and offline features, a Pythonic developer experience, and cost visibility in one place as its selection criteria.
VALID Systems cuts deposit loss rates 7%-10% after rearchitecting on Snowflake
VALID Systems guarantees over $6 billion in immediately available funds each month across more than 425 million unique accounts. It covers teller, ATM and mobile channels. After its legacy solution reached its ceiling, VALID moved features, training data, models and inference onto Snowflake. It kept the same API contract with zero egress.
VALID's models now optimize for the dollar value of each decision rather than row-level accuracy. The company says a small gain on the highest-risk transactions outweighs a larger gain spread evenly. Every tuning run is logged and auditable, so compliance documentation is a byproduct of the process.
VALID reports a 7%-10% drop in deposit loss rates and detection of 15%-17% more bank fraud losses.
Legacy batch pipelines miss millisecond SLAs, Snowflake argues
Snowflake says batch pipelines and static rules engines cannot evaluate live transactions against fresh behavioral data within millisecond SLAs. It adds that legacy systems often send inference calls to external platforms, creating data egress and governance risk for regulated institutions.
Online Feature Store serves features in under 10 ms and models in under 100 ms
The Online Feature Store is built on Snowflake Postgres. It serves risk scores, activity patterns and behavioral features to models at prediction time. These workloads require thousands of predictions per second.
Trained models run as managed inference services on Snowpark Container Services. Any model in the Snowflake Model Registry can be deployed to a dedicated REST endpoint with a single API call. Model serving takes under 100 ms. The service fetches the latest features automatically, so a fraud scoring decision, from feature lookup to prediction, completes in one network round trip.
Stream feature views deliver end-to-end freshness under 200 ms p50
Stream feature views ingest transactions, logins and device events through a REST ingestion API. Events are written synchronously to a managed Postgres serving layer. The store maintains running time-windowed aggregations, such as total spend in the last 48 hours or transaction count in the last hour, instead of rescanning raw event history.
Real-time feature views run a Python function at inference time. It combines per-request inputs, such as transaction amount or device fingerprint, with precomputed features. A model can, for example, compare the current transaction amount to a customer's rolling average spend to flag an anomaly as it occurs.