Bud Financial

Bud Financial serves as the customer-intelligence layer for banks, credit unions, and fintechs—a segment where data enrichment quality directly determines product trust and regulatory standing. Raw payment transactions arrive as opaque strings with no merchant context, category, or behavioral signal. Bud's mandate was to classify each transaction (e.g., Food → Coffee → Coffee shops), attach merchant identity and geolocation, and detect recurring patterns before the customer's payment notification fires. The binding constraint was the credit card authorization loop: every enrichment decision had to complete within 5 milliseconds. At 500 million monthly transactions for a single client, any mislabeled payment—an incorrect category, a false fraud trigger—cascades into disputes and compliance exposure, making precision non-negotiable from the outset.

Bud built a modular, real-time enrichment pipeline anchored by proprietary machine learning and predictive analytics models trained and maintained entirely in-house. The pipeline ingests both real-time payment streams and batch banking feeds, runs classification and merchant mapping in milliseconds, and returns enriched data within the authorization window. For the transactional data store, Bud migrated from self-managed Apache Cassandra to Astra DB (part of IBM watsonx.data), gaining a fully managed, globally distributed service capable of handling unpredictable volume spikes while satisfying per-client data-residency and hybrid deployment requirements. Large foundation models were deliberately excluded from the real-time path to eliminate hallucination risk and preserve explainability—a hard requirement in regulated banking. Enriched transactions stream to Google BigQuery for clients requiring cross-customer analytics, cleanly separating transactional and analytical workloads. A human-in-the-loop labeling process continuously verifies ambiguous or emerging transaction types to keep models calibrated as new payment categories evolve.

The enrichment engine delivers >97% classification accuracy across a global merchant network of 20 million merchants and 30 million locations, processing 500M+ transactions monthly for Bud's largest client—a retail bank serving more than 10 million users—with sub-5ms latency throughout. HSBC adopted the platform in 2019 and subsequently became an investor, validating the architecture at global banking scale.

• Engage: Increases digital-channel adoption by giving retail customers a clearer, more trusted view of their spending behavior.
• Assess: Expands credit access for thin-file borrowers by evaluating real-time affordability and spending patterns rather than relying solely on traditional credit history.

• Human-in-the-loop verification sustains accuracy over time—continuously reviewing ambiguous categories (including emerging types like BNPL and ridesharing) is what keeps classification above 97% in production.
• Exclude foundation models from latency-critical loops; purpose-built ML preserves both the explainability regulators require and the sub-millisecond response times payment infrastructure demands.
• Managed database services reduce operational overhead at banking scale—migrating from self-managed Cassandra to a fully managed service freed engineering capacity while satisfying per-client data-residency contracts.
• Separate transactional and analytical workloads by design; real-time enrichment and cross-customer analytics have incompatible latency profiles and belong on distinct infrastructure.