Patents
A selection of granted GenAI patents spanning AI compliance validation, agentic AI governance, and multi-model monitoring. The full list — including additional co-patents not shown here — is at Google Patents.
Lead Patents — Regulatory Compliance
Citibank, N.A. · US12111754B1 · Granted October 8, 2024
A system that evaluates the compliance of an AI application with predefined guidelines. The technology obtains a set of guidelines defining operation boundaries of the AI application and constructs test cases associated with each guideline — each comprising a prompt, an expected outcome, and an expected explanation. It supplies the prompts to the AI application, receives case-specific outcomes and explanations, and compares them with expected results. A compliance indicator is generated based on the evaluation, indicating the degree of compliance of the AI application with the guidelines.
Citibank, N.A. · US12361335B1 · Granted July 15, 2025
A system that evaluates the compliance of an AI application with predefined vector constraints. Multiple specialised models — trained to identify specific types of non-compliance — analyse the representation of attributes within AI-generated responses, identify vector representations of alphanumeric characters by assessing proximate locations, frequency, and associations, and determine indicators of vector alignment by measuring differences in the direction or magnitude of vector representations against predetermined characters.
Co-Patents — AI Applications
Citibank, N.A. · US12198030B1 · Granted January 14, 2025
The technology evaluates the compliance of an AI application with predefined vector constraints, employing multiple specialised models trained to identify specific types of non-compliance within AI-generated responses. Models analyse the representation of attributes within responses, identify vector representations of alphanumeric characters by assessing proximate locations, frequency, and associations with other characters, and determine indicators of vector alignment by measuring differences in direction or magnitude of vector representations against predetermined characters.
Citibank, N.A. · US12299140B2 · Granted May 13, 2025
A multi-model superstructure that dynamically routes the artifacts of a first set of models to one or more models of a second set for evaluation against assessment metrics. The approach mitigates overfitting and security vulnerabilities by frequently rotating monitoring models rather than relying on static defence mechanisms, making it harder for adversarial actors to exploit predictable vulnerabilities in AI systems.
Citibank, N.A. · US12346820B1 · Granted July 1, 2025
Systems that receive regulatory definitions establishing boundaries for expected AI use cases, paired with operational data. An initial AI model constructs observed use cases from the operational data; a second AI model correlates regulatory definitions and use case attributes to an appropriate risk classification. The system identifies compliance shortfalls within each observed use case and produces recommended remediation actions, enabling continuous alignment of AI deployments to regulatory requirements.
Citibank, N.A. · US12450494B1 · Granted October 21, 2025
A proactive validation layer that continuously monitors and evaluates the actions of autonomous agents using generative AI models. The system receives constraints and operational data defining agent parameters, dynamically evaluates agent behaviours for compliance with predefined boundaries and regulatory standards, maps agents to risk categories, generates expected actions, identifies gaps between expected and actual outputs, and modifies proposed actions through AI-driven analysis — enabling agents to execute refined actions while preventing unauthorised behaviour and maintaining audit trails for transparency.
Citibank, N.A. · US12505352 · Granted December 23, 2025
The system employs generative AI models to identify and remediate compliance gaps within AI use cases. It creates an inventory of observed model applications through regulatory definitions or existing records, classifies each into risk categories per applicable regulations, identifies compliance shortfalls, and generates action plans addressing gaps — enabling continuous lifecycle evaluation and automated documentation aggregation for regulatory adherence.
Earlier Patents
WO2008152530A1 · Filed April 13, 2008 · Published December 18, 2008
A playpens system offering shape adaptability to suit various floor spaces. The design can be folded and collapsed for storage and transport, and accommodates optional add-on modules — gates, playhouses, toy boxes — for upgradability. Frames utilise hollow PVC, soft wood, or metal construction with netted fabric panels, connected via specialised L-Sockets and Frame Connectors for assembly into various geometric configurations. Co-invented with Basil K.P.B.I. de Silva.