Sardine AI Labs
Sardine AI Labs
Advancing the frontier of intelligence to fight financial crime
Sardine AI Labs is an applied research group building AI to tackle some of the hardest problems in financial crime.
Grounded in Sardine’s network of device, identity, behavioral, and transaction data, we are developing models that can understand complex financial behavior and detect attacks they were never explicitly trained to find.
Understanding the language of financial crime
Our research tackles open problems in sequential modeling, transfer learning, adversarial robustness, and explainability.
We test each advance against the demands of production risk systems, including detection performance, cross-institution generalization, and low-latency inference. Our goal is to turn advances in AI into measurable protection against financial crime.
Purpose built for risk
Trained on our large repository of actual fraudulent transactions, financial crime data, and observed outcomes.
Engineered for production
Model accuracy is measured alongside latency, governance, explainability, and the requirements for human oversight.
Transparent research
We publish what worked, where it failed, and the questions the evidence cannot answer yet.
Learnings from our first foundation model for issuing risk
Our first payments foundation model learned from unlabeled transaction sequences, then supplied new features to the issuing fraud model already in production.
more fraud caught at the same share of transactions flagged for action
improvement in overall ranking quality, measured by AUC-PR
improvement for a client population held out of pretraining
Recall by action depth
- 0.05%7.2%
- 0.10%13.4%
- 0.50%35.5%
- 1.00%47.5%
- 2.00%58.7%
Action depth is the share of scored traffic selected for decline, freeze, or review.
Learn first, classify second
Tokenize merchant, amount, time, geography, channel, and terminal.
Pretrain on 10 months of unlabeled histories.
Represent each history with an eight-layer transformer.
Combine the features with the existing fraud model.
Built on the fastest-growing dataset in financial crime
Our research starts with the breadth of Sardine’s network and goes deep into the sequence of each customer’s financial behavior.
Data scope and schema
Our broad data consortium captures how financial behavior evolves over time across sessions, transactions, and customer profiles. Here’s a look at the dataset that powers our Issuing Risk Foundation Model.
Data
Card-issuing histories from 13 issuers across consumer, SMB, crypto-card, and neobank programs. Coverage runs from January 2024 to July 2026, with up to 31 months per issuer.
Learning Unit
Each cardholder’s complete history becomes a single sequence, reaching approximately 2,300 transactions.
Schema
Each transaction contains 14 fixed fields spanning merchant, MCC, amount, timing, geography, entry mode, presence context, terminal information, and card balance.
Vocabulary
Approximately 72,000 tokens. The top 50,000 merchants receive dedicated tokens, the long tail maps to 20,000 stable hash buckets, and numeric values are quantile-binned.
Model
A Llama-style decoder with approximately 300 million parameters, trained through next-token prediction. It produces a 640-dimensional embedding for every transaction.
Research grants up to $50,000
We are awarding up to five (5) research grants to independent researchers that can help us tackle our research topics.
Research areas we’re funding
Cross-issuer transfer
How can representations learned across participating issuers improve fraud detection for an institution the model has never seen? In early testing on a fully held-out issuer, embeddings increased AUC-PR from approximately 0.05-0.07 to 0.14-0.21 compared with a strong hand-crafted-feature baseline.
Multi-stream event modeling
How can device intelligence, behavioral biometrics, onboarding and KYC, logins, card transactions, ACH and transfers, and disputes be modeled as a single customer sequence? Research may address shared tokenization, relationships between events, and the value of sparse signals relative to dense transaction activity.
Scaling laws for financial models
How does performance change as the number of issuers, length of customer histories, and model size increase? Research should help establish whether consistent scaling laws exist for financial event models.
Cold-start learning
How much history does a model need to produce useful representations for a new cardholder or issuer? Research should establish when embeddings begin to outperform aggregation-based features.
Applications beyond fraud
Can representations learned from financial activity transfer to other predictive tasks? Potential applications include credit risk, churn prediction, spend forecasting, and merchant categorization.
Recent publications
Building a foundation risk model
Real-world learnings from training an AI model for issuing.
We trained a language model and it caught 35% more fraud
Financial data is ordered, relational, and sparse. General-purpose models were not built for it.
Welcome to the Agentic Frontier for Risk Ops
A plain-language guide to turning payment histories into something a transformer can learn from.
Why your fraud AI works better inside your fraud platform
What might happen when device, identity, behavior, and payments share a representation.