Abstract: Certain aspects of the present disclosure provide techniques and systems for screening chat attachments. A chat attachment screening system monitors a chat window of a first computing device associated with a first user during an interaction session between the first user and a second user. An upload of an attachment is detected based on the monitoring. Access to the attachment from a second computing device associated with the second user is blocked, in response to detecting the upload. Content from the attachment is identified and extracted. A type of the attachment is determined based on the content. A determination is made as to whether the second user is authorized to access the type of the attachment. An indication of the determination is presented on at least one of the first computing device or the second computing device during the interaction session.

Abstract: Systems and methods are disclosed for detecting hallucinations in large language models (LLMs). An example method includes receiving a first prompt for submission to the first LLM, generating, using the first LLM, a plurality of semantically equivalent prompts to the first prompt, generating, using the first LLM, a first response to the first prompt and a plurality of second responses to the plurality of semantically equivalent prompts, generating, using a second LLM, a plurality of third responses to the semantically equivalent prompts, generating a semantic consistency score for the first response based at least in part on the first prompt, the plurality of semantically equivalent prompts, the plurality of second responses, and the plurality of third responses, and determining whether or not the first response is an accurate response to the first prompt based at least in part on the semantic consistency score.

Abstract: Systems and methods for training an encoder-decoder model are disclosed.

Abstract: A first large language model (LLM) instance may be instructed to request data while being prevented from performing calculations using the data. A second LLM instance may be instructed to provide a response to the request for data based on a known complete data set. The response may be translated into a machine-readable response in a format configured for processing by a calculation engine. The calculation engine may process the machine-readable response, thereby generating a calculation engine output. A mismatch between the calculation engine output and a known result obtained using the known complete data set may be identified, and the instruction to the first LLM may be modified in response.

Abstract: This disclosure relates to systems and methods for identifying fraudulent email addresses associated with an electronic payment service. In some implementations, a computing device receives an email with a prefix having a number of characters and characterized by a prefix length indicative of the number of characters in the prefix. The computing device identifies each of a number of bigrams is identified within the prefix, and determines a row and column distance for each bigram between two consecutive characters of the bigram as positioned on a keyboard. The computing device calculates a Euclidean distance between the two consecutive characters of the bigram based on the row and column distances, and determines a normalized distance based on the prefix length and an average of the Euclidean distances calculated for the number of bigrams in the prefix. The normalized distance is compared with a value to classify the email as suspicious or as not suspicious.

Abstract: Certain aspects of the disclosure provide a method for managing users, the method comprising: obtaining a first set of visitor identification records; identifying a subset of outlier visitor identification records within the first set of visitor identification records; creating a second set of visitor identification records including all visitor identification records from the first set of visitor identification records other than the subset of outlier visitor identification records; creating a third set of visitor identification records by applying a first filtering scheme to the second set of visitor identification records; creating a fourth set of visitor identification records by applying a second filtering scheme to the second set of visitor identification records; generating one or more stitched visitor IDs by stitching visitor identification records in the fourth set of visitor identification records; and determining a stitching accuracy probability for each of the one or more stitched visitor IDs.

Abstract: At least one large language model (LLM) may be instructed to request data from a user while being prevented from performing calculations using the data. A user-generated response to the request for data, including at least a portion of the data, may be received. The user-generated response may be translated into a machine-readable response in a format configured for processing by a calculation engine. The calculation engine may process the machine-readable response, thereby generating a calculation engine output.

Abstract: Aspects of the present disclosure provide techniques for classifying a trip. Embodiments include receiving, from a plurality of users, a plurality of historical trip records. Each of the plurality of historical trip records may comprise one or more historical trip attributes and historical classification information. Embodiments include training a predictive model, using the plurality of historical trip records, to classify trips based on trip records. Training the predictive model may comprise determining a plurality of hot spots based on the historical trip records, each of the plurality of hot spots comprising a region encompassing one or more locations, and associating, in the predictive model, the plurality of hot spots with historical classification information. Embodiments include receiving, from a user, a new trip record comprising a plurality of trip attributes related to a trip and using the predictive model to predict a classification for the trip based on the trip record.

Abstract: Certain aspects of the present disclosure provide techniques for processing transactions in a computing system. An example method generally includes receiving a request to perform an operation with respect to an object included in the request. A system identifies an archetype defining properties of the object included in the request. Based on the identified archetype, the system identifies data repositories to commit data to in order to perform the requested operation and rules for performing the operation with respect to the object. One or more actions are executed against the identified data repositories according to the identified rules.

Abstract: Augmented Denoising Diffusion Implicit Models (“DDIMs”) using a latent trajectory optimization process can be used for image generation and manipulation using text input and one or more source images to create an output image. Noise bias and textual bias inherent in the model representing the image and text input is corrected by correcting trajectories previously determined by the model at each step of a diffusion inversion process by iterating multiple starts the trajectories to find determine augmented trajectories that minimizes loss at each step. The trajectories can be used to determine an augmented noise vector, enabling use of an augmented DDIM and resulting in more accurate, stable, and responsive text-based image manipulation.

Abstract: A processor may receive user information for a request payload from an external device and data describing a plurality of user interface (UI) elements configured to be presented in a UI of the external device. The processor may select a machine learning (ML) model from a plurality of ML models using a contextual bandit ML model that is trained based on the user information. The processor determines at least one recommended user interface (UI) element with a selected ML model, based on the user information and the data describing the plurality of UI elements. The at least one recommended UI element may be presented in the UI of the external device. The processor may receive event data indicating a user interaction with the at least one recommended UI element in the UI of the external device. The contextual bandit ML model may be re-trained based on the event data.

Abstract: An orchestration layer for execution user defined automation workflows. The orchestration layer may include multiple process instances that host user defined automation workflows that automate processes or tasks. To improve system performance and reduce operating costs, the user defined automation workflows are deployed to the orchestration layer in a standard format that standardizes the user defined workflow configurations. The orchestration layer may also dynamically scale the computational resources allocated to teach process instance based on the properties of each user defined automation workflow.

Abstract: A method includes obtaining matches between target records in a target dataset and a reference records in a reference dataset, each match of the matches comprising a corresponding confidence level of the match, categorizing the target records into review level categories according to the corresponding confidence level, and presenting a graphical user interface (GUI). The GUI includes a first section for a first review level category showing a first subset of the target records assigned to the first review level category, the first subset comprising target records related, in the GUI, to at least one matching reference record. The GUI includes a second section for a second review level category, wherein the second section shows a second subset of the target records assigned to the second review level category, the second subset comprising target records related, in the GUI, to at least one matching reference record.

Abstract: Systems and methods are provided for generating recommendations using contextual bandit models with non-linear oracles.

Abstract: A method for detecting fraudulent financial transactions in information technology networks involves obtaining a multitude of features associated with a financial transaction conducted over an information technology network by an unknown transaction party. The multitude of features includes clickstream data obtained from the unknown transaction party. The clickstream data is associated with data of the financial transaction being entered by the unknown transaction party. The method further involves obtaining a first fraud indicator using a machine learning classifier operating on the multitude of features, obtaining a second fraud indicator using a rule-based classifier operating on the multitude of features, obtaining a fraud prediction for the financial transaction, using the first fraud indicator and the second fraud indicator, and taking an action, in response to the fraud prediction.

Abstract: Certain aspects of the disclosure provide a method, comprising: processing input data with an ensemble of nonlinear machine learning models; generating a sparse high-dimensional embedding based on one or more leaf nodes of each nonlinear machine learning model in the ensemble of nonlinear machine learning models; projecting the high-dimensional embedding into a lower-dimensional embedding, wherein the lower-dimensional embedding is less sparse than the high-dimensional embedding; processing the lower-dimensional embedding with a linear machine learning model to generate a binary class prediction; determining a confidence for the binary class prediction; and outputting: the binary class prediction if the confidence is greater than or equal to a threshold; or a flipped binary class prediction if the confidence is lower than the threshold.

Abstract: Systems and methods for enriching raw user text with a database to identify relevant context, wherein generated prompts provide responses to user queries is provided. A method includes receiving a query, wherein the query comprises the raw text, creating a first embedding based on the query, retrieving a plurality of other embeddings, wherein the plurality of other embeddings are complementary to the first embedding, creating a contextual prompt including context from at least one of the plurality of other embeddings, processing the contextual prompt using a trained machine learning model, thereby generating a response to the query, and causing the response to be displayed by a display device.

Abstract: A method classifies feedback from transcripts. The method includes receiving an utterance from a transcript from a communication session and processing the utterance with a classifier model to identify a topic label for the utterance. The classifier model is trained to identify topic labels for training utterances. The topic labels correspond to topics of clusters of the training utterances. The training utterances are selected using attention values for the training utterances and clustered using encoder values for the utterances. The method further includes routing the communication session using the topic label for the utterance.