DETAILED ACTION
This communication is in response to the application filed 6/29/2023. Claims 1-20 are pending in the application.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 6/29/2023, 8/2/2024, 9/9/2024, 9/12/2024, 10/2/2024, 10/11/2024, 12/6/2024, 1/13/2025, 2/4/2025, 4/8/2025, 4/21/2025, 5/21/2025, 6/20/2025, 8/12/2025, 9/9/2025, 9/17/2025, 10/13/2025, 4/6/2026, 5/8/2026, 6/17/2026 are in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-6, 10-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Zhang (US 12481517) in view of Sankaranarayanan (US 20230013479).
As per claims 1, 10, 16, Zhang (US 12481517) teaches
a method of managing a data pipeline, the method comprising: making a first identification that a first translation schema has a first performance score that falls below a performance score threshold (col. 2, last para: a translation layer configured to standardize input and output schemas for interacting with the plurality of AI agents. The model orchestration subsystem uses this translation layer to send one or more prompts to the identified AI agents and receive the returned data from those agents; col. 4:36-43; col. 20:21-28: if performance lags behind predefined thresholds, hyperparameter tuning routines or architecture search algorithms are triggered automatically. These computational strategies guarantee that the final model operates reliably under diverse conditions),
the first translation schema being intended to remediate a change in a system of representation of information conveyed by data obtained from a data source and the data source comprising a stochastic element that influences the data (col. 16:31-43: the system periodically retrains and fine-tunes the LLM on updated domain-specific datasets. This involves automatically ingesting new training samples, augmenting them to increase coverage, and running iterative optimization steps such as stochastic gradient descent on large distributed training clusters. Performance metrics, stored in a model registry, guide the automatic selection of model checkpoints that exhibit improved accuracy. In this way, the model continuously refines its parameters-without human intervention-ensuring that it adapts to emerging linguistic patterns, product lines, or seasonal trends, ultimately delivering precise intent classification results at scale; col. 4:36-42: the system includes a translation layer that standardizes input and output schemas for interacting with multiple AI agents. By using a unified schema, the translation layer ensures interoperability across AI agents that may have differing input requirements and response formats);
obtaining, in response to the first identification, a second translation schema based, at least in part, on synthetic data from a synthetic data source, the synthetic data source being intended to generalize operation of the data source; the synthetic data source excluding the stochastic element so that the synthetic data is not influenced by the stochastic element (col. 6:19-32: the training process for the user intent LLM may start with the acquisition of a dataset comprising a diverse range of user queries paired with corresponding ground truth intent labels. This dataset is designed to reflect the specific domain of application (e.g., coffee shops, fashion retail, pizza delivery), capturing a variety of query formats, phrasings, and contextual variations to ensure that the LLM can effectively generalize across different scenarios. Possible sources for this training data include logs from customer interactions, such as chat transcripts, support tickets, email communications, and voice-to-text transcriptions from call centers. Additionally, publicly available datasets, such as e-commerce question-answering datasets, or synthetically generated data can supplement real-world examples; col. 7:last para.: the system includes a translation layer that standardizes input and output schemas for interacting with multiple AI agents. By using a unified schema, the translation layer ensures interoperability across AI agents that may have differing input requirements and response formats; col. 16:31-43: the system periodically retrains and fine-tunes the LLM on updated domain-specific datasets. This involves automatically ingesting new training samples, augmenting them to increase coverage, and running iterative optimization steps such as stochastic gradient descent on large distributed training clusters. Performance metrics, stored in a model registry, guide the automatic selection of model checkpoints that exhibit improved accuracy. In this way, the model continuously refines its parameters-without human intervention-ensuring that it adapts to emerging linguistic patterns, product lines, or seasonal trends, ultimately delivering precise intent classification results at scale);
making a first determination regarding whether the second translation schema has a second performance score that meets the performance score threshold; in an instance of the first determination in which the second translation schema has the second performance score that meets the performance score threshold: performing an action set to implement the second translation schema in the data pipeline (col. 3:35-41: the system selects the returned data with the highest evaluation score as the final response. If multiple AI agents return responses, the system combines them by removing redundant information, aggregating non-overlapping portions, and synthesizing a unified response using the LLM; col. 4:36-43; col. 10, last paragraph: if multiple AI agents return responses, the orchestration subsystem assigns an evaluation score to each response based on these criteria. A response with a higher confidence score and closer alignment to the user intent, for instance, would be prioritized over one with a lower confidence score or incomplete information. The system then selects the response with the highest evaluation score as the final output; col. 12, last para: the model's development begins with the creation of a labeled dataset, where each user request is paired with a specific context category—such as “real-time interaction,” “delayed task,” or “refinement task.” Constructing this dataset involves automated data ingestion pipelines that collect user interaction logs, system request histories, and meta-information (e.g., timestamps, input modalities), ensuring the dataset accurately reflects real-world conditions. To maintain data quality, the dataset undergoes a computer-implemented preprocessing pipeline).
Zhang does not explicitly teach a synthetic data source.
Sankaranarayanan teaches at para. 49: the datasets may be in different forms, such as files, tables, databases, autonomous databases, object stores…; para. 149: facilitate data cataloging and synthetic dataset generation; figs. 2, 7-8: synthetic missing dataset generation model, synthetic replacement dataset generation model; para. 69: the inclusion of the synthetic data in the synthetic datasets makes the synthetic datasets better training data candidates than their corresponding original datasets, thus improving the accuracy and performance of the models trained using the synthetic datasets.
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Zhang to include the synthetic data source of Sankaranarayanan in order to effectively improving the accuracy and performance of the models trained using the synthetic datasets since by generating multiple synthetic datasets automatically, the data catalog system enables a vast amount of diverse training data to be made available for use by data scientists. The diversity of the synthetic dataset that is generated enables it to be used for training models for various different applications. Further, the inclusion of the synthetic data in the synthetic datasets make the synthetic datasets better training data candidates than their corresponding original datasets – See Sankaranarayanan, para. 46.
As per claims 2, 11, 17, Zhang teaches
obtaining the first translation schema (col. 2, last para.: the system also includes a translation layer configured to standardize input and output schemas for interacting with the plurality of AI agents; col. 4:36-40: to facilitate seamless integration with existing enterprise systems, such as POS systems, back-end databases, and warehouse management systems, the system includes a translation layer that standardizes input and output schemas for interacting with multiple AI agents).
Zhang does not explicitly teach prior to making the first identification: making a second determination regarding whether the data comprises anomalous data, the anomalous data indicating the change in the system of representation of information; and in an instance of the second determination in which the data comprises the anomalous data.
Sankaranarayanan teaches said limitations at para. 93: profiler subsystem may provide insights into various column and row level statistics, detect patterns, anomalies, and relationships in the data. The profiling may be performed on a continuous basis as new original datasets are cataloged and new synthetic datasets are created. Various different types of profiling may be performed by profiler subsystem including attribute profiling, structural profiling, and schema profiling; para. 96.
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Zhang to include determination regarding whether the data comprises anomalous data, the anomalous data indicating the change in the system of representation of information of Sankaranarayanan in order to effectively improving the accuracy and performance of the models trained using the synthetic datasets since the inclusion of the synthetic data in the synthetic datasets makes the synthetic datasets better training data candidates than their corresponding original datasets – See para. 69.
Specification, para. 128 discloses “Obtaining the degree of anomalousness for the data may include: (i) performing a cluster analysis using the data, (ii) performing an isolation forest process using unsupervised machine learning, (iii) performing a statistical analysis to compare each element of the data to historic data trends (e.g., via determining a number of standard deviations away from the historic data mean each element of the data is), and/or (iv) other methods”.
As per claims 3, 12, 18, Zhang teaches
wherein obtaining the first translation schema comprises: obtaining first historic data, the first historic data being previously provided to one or more downstream consumers and the first historic data being based on a first system of representation of information (col.3:49-58: to classify the user request, the system applies a machine agents, the orchestration subsystem identifies candidate AI
learning-based intent classification model that analyzes textual content, input modality, temporal patterns, and user interaction history; col. 10:20-36; the orchestration subsystem analyzes the extracted user intent in conjunction with various real time
model metrics-such as latency, cost, availability, or historical performance-to determine which AI agents are best suited for the task at hand; col. 15:17-36);
issuing a first request for the first historic data from the data source to obtain an updated instance of the first historic data, the updated instance of the first historic data being based on a second system of representation of information (col. 7, last paragraph: to maintain relevance over time, the system may integrate
mechanisms for real-time updates. As new user intents emerge or existing ones evolve, the dataset and embeddings are updated to reflect these changes. Regular model retraining ensures that the LLM adapts to shifting domain requirements while preserving backward compatibility through robust versioning controls. This dynamic update mechanism enables the LLM to remain accurate, responsive, and effective, even in rapidly changing environments; col. 12:36-67).
Sankaranarayanan also teaches at para. 87-90: catalog generator subsystem is configured to generate and/or update original catalogs for one or more customers. An original data catalog generated by catalog generator subsystem for a customer may store information related to the original datasets for the customer that are accessed by data access subsystem. After an original data catalog has been generated for a customer, as additional datasets are cataloged for that customer, the original data catalog is updated by catalog generator subsystem for those additional original datasets. When a synthetic dataset is generated for an original dataset, the catalog generator subsystem is also responsible for updating the catalog information for that original dataset in the original data catalog to include information about the synthetic dataset. This enables users accessing the original data catalog to, for an original dataset, have knowledge of any synthetic datasets that have been generated based upon that original dataset.
As per claims 4, 13, 19, Zhang teaches
wherein making the first identification comprises: obtaining the first performance score, the first performance score indicating a degree to which the first translation schema successfully remediates the change in the system of representation of information; and comparing the first performance score to the performance score threshold (col. 3:28-41: performance metrics such as precision, recall, and FI-score are calculated based on the predicted user intent and the retrieved intent examples. The LLM is iteratively refined based on these performance metrics. The system selects the returned data with the highest evaluation score as the final
response; col. 7:26-37: the retrieved examples are incorporated into the LLM's input to improve the contextual relevance of its predictions. The system evaluates the performance of the LLM by calculating metrics such as precision, recall, and Fl-score,
which measure the model's ability to extract accurate intents. These performance metrics guide an iterative refinement process, where the model's parameters are adjusted to address identified weaknesses, further enhancing its accuracy
and robustness).
As per claims 5, 14, 20, Zhang teaches
wherein an influence of the stochastic element on the data negatively impacts the first performance score (col. 3:28-52: performance metrics such as precision, recall, and FI-score are calculated based on the predicted user intent and the retrieved intent
examples. The LLM is iteratively refined based on these performance metrics. The system selects the returned data with the highest evaluation score as the final response; col. 16:31-43: as user behavior evolves, the system periodically retrains and fine-tunes the LLM on updated domain-specific data sets. This involves automatically ingesting new training samples, augmenting them to increase coverage, and running iterative optimization steps such as stochastic gradient descent on large distributed training clusters).
Sankaranarayanan also teaches said limitation at para. 45: the accuracy, and thus the usability, of these models depends upon the quality of the training data. It is preferred that the training dataset is as complete as possible—training with a dataset that has holes leads to inaccurate models, which lead to bad outcomes. Even if a dataset is complete, many times the dataset contains restricted data that has to be removed from the dataset before the dataset can be made available for training. However, removing this restricted data again creates holes in the dataset leading to inaccurate models, as described above. In some instances, the entire dataset is marked as restricted leading to the entire dataset being unavailable for training. Given the nature of enterprise datasets, a large number of the datasets are not appropriate for ML training because of the problems described above. As a result, even though historical data is available for training purposes, it cannot be use. Thus, negatively impacts the first performance score.
As per claims 6, 15, Zhang teaches
wherein obtaining the second translation schema comprises: obtaining the first historic data (col. 2, last para.: The system also includes a translation layer configured to standardize input and output schemas for interacting with the plurality of AI agents; col. 3:49-52: To classify the user request, the system applies a machine agents, the orchestration subsystem identifies candidate AI learning-based intent classification model that analyzes textual content, input modality, temporal patterns, and user
interaction history).
Zhang does not explicitly teach issuing a second request for the first historic data from the synthetic data source to obtain the synthetic data, the synthetic data being based on the second system of representation of information; mapping portions of the synthetic data to corresponding portions of the first historic data to identify a second relationship between the first system of representation of information and the second system of representation of information; and obtaining the second translation schema based on the second relationship.
Sankaranarayanan teaches said limitations in fig. 2: identify and access the original dataset from its source data/first historic data; figs 7-8; para. 141-142: as depicted in FIG. 7, an original dataset is a row-and-column dataset comprising columnar values corresponding to values for a “NAME,” “SOCIAL SECURITY NUMBER (SSN),” “LOCATION (LOC.),” and “TELEPHONE NUMBER (TEL.).” Original dataset 700 may comprise missing data 710. For example, as depicted in FIG. 7, a subset of data values in the “LOC.” column are not present or are null. In response to determining that original dataset has one or more missing values, original dataset may be input into synthetic missing dataset generation model. Synthetic missing dataset generation model may be a machine-learning model such as a convolutional neural network model which will take, as input, the original dataset to produce synthetic data values for a new synthetic dataset. Specifically, synthetic missing dataset generation model may output synthetic data values to populate new synthetic dataset; para. 146: the mapping may cause generation of new synthetic data values which can be used to populate new synthetic dataset. For example, the values in new synthetic dataset appear somewhat similar to values in original dataset but with some differences.
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Zhang to include mapping portions of the synthetic data to corresponding portions of the first historic data of Sankaranarayanan in order to effectively improving the accuracy and performance of the models trained using the synthetic datasets since the inclusion of the synthetic data in the synthetic datasets makes the synthetic datasets better training data candidates than their corresponding original datasets – See para. 69.
Claim(s) 7-9 is/are rejected under 35 U.S.C. 103 as being unpatentable over Zhang (US 12481517) in view of Sankaranarayanan (US 20230013479) and further in view of Yeri (US 20240411838).
As per claim 7, Zhang, Sankaranarayanan do not explicitly teach claim 7.
Yeri teaches
wherein the synthetic data source comprises one selected from a list consisting of: a digital twin of the data source; and an inference model trained to generalize the operation of the data source (para. 68: the prediction generation circuitry may use an inference machine learning model to process the stimulus corresponding to the synthetic stimulus and the feature set and/or user data associated with the digital twin; para. 78).
Thus, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to modify the teachings of Zhang, Sankaranarayanan to include a digital twin of the data source and an inference model trained to generalize the operation of the data source of Yeri in order to identify and use a digital twin that is similar to target user and has experienced a stimulus corresponding to the synthetic stimulus may provide real-world insights into user behavior and furthermore, may be less computationally expensive by leveraging existing user data – See Yeri, para. 4.
As per claim 8, Zhang teaches
wherein making the first determination comprises: obtaining the second performance score, the second performance score indicating a degree to which the second translation schema successfully remediates the change in the system of representation of information; and comparing the second performance score to the performance score threshold (col. 3:28-41: performance metrics such as precision, recall, and FI-score are calculated based on the predicted user intent and the retrieved intent examples. The LLM is iteratively refined based on these performance metrics. The system selects the returned data with the highest evaluation score as the final response; col. 4:36-42: the system includes a translation layer that standardizes input and output schemas for interacting with multiple AI agents; col. 7:26-37: the retrieved examples are incorporated into the LLM's input to improve the contextual relevance of its predictions. The system evaluates the performance of the LLM by calculating metrics such as precision, recall, and Fl-score, which measure the model's ability to extract accurate intents. These performance metrics guide an iterative refinement process, where the model's parameters are adjusted to address identified weaknesses, further enhancing its accuracy and robustness; col. 16:31-43: the system periodically retrains and fine tunes the LLM on updated domain-specific datasets. This involves automatically ingesting new training samples, augmenting them to increase coverage, and running iterative optimization steps such as stochastic gradient descent on large distributed training clusters. Performance metrics/scores, stored in a model registry, guide the automatic selection of model checkpoints that exhibit improved accuracy. In this way, the model continuously refines its parameters without human intervention, ensuring that it adapts to emerging linguistic patterns, product lines, or seasonal trends, ultimately delivering precise intent classification results at scale).
As per claims 9, Zhang teaches
wherein performing the action set comprises: obtaining a translation layer for the data pipeline, the translation layer being adapted to initiate implementation of the second translation schema when future instances of data based on the second system of representation of information are identified (col. 2, last para.: the system also includes a translation layer configured to standardize input and output schemas for interacting with the plurality of AI agents; col. 6:56-61; col. 7, last para.: Regular model retraining ensures that the LLM adapts to shifting domain requirements while preserving backward compatibility through robust versioning controls. This dynamic update mechanism enables the LLM to remain accurate, responsive, and effective, even in rapidly changing environments; col. 10, last para. A response with a higher confidence score and closer alignment to the user intent, for instance, would be prioritized over one with a lower confidence score or incomplete information. The system then selects the response with the highest evaluation score as the final output).
Sankaranarayanan also teaches said limitations in figs 7-8, para. 141-142: as depicted in FIG. 7, an original dataset is a row-and-column dataset comprising columnar values corresponding to values for a “NAME,” “SOCIAL SECURITY NUMBER (SSN),” “LOCATION (LOC.),” and “TELEPHONE NUMBER (TEL.).” Original dataset may comprise missing data. For example, as depicted in FIG. 7, a subset of data values in the “LOC.” column are not present or are null. In response to determining that original dataset has one or more missing values, original dataset may be input into synthetic missing dataset generation model. Synthetic missing dataset generation model may be a machine-learning model such as a convolutional neural network model which will take, as input, the original dataset to produce synthetic data values for a new synthetic dataset. Specifically, synthetic missing dataset generation model may output synthetic data values to populate new synthetic dataset; para. 146: the mapping may cause generation of new synthetic data values which can be used to populate new synthetic dataset. For example, the values in new synthetic dataset appear somewhat similar to values in original dataset but with some differences.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Lang (US 20240202405) teaches at para. 99: inference—using a trained AI model to make predictions on new data or generate new data.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to LINH BLACK whose telephone number is (571)272-4106. The examiner can normally be reached 9AM-5PM EST M-F.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Tony Mahmoudi can be reached at 571-272-4078. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/LINH BLACK/Examiner, Art Unit 2163 5/20/2026
/TONY MAHMOUDI/Supervisory Patent Examiner, Art Unit 2163
Prosecution Insights