# US 2026/0065093 A1 · Artificial intelligence hybrid distributed inference system

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0065093 A1
- Document ID: US20260065093A1
- Status: Published application
- Kind: A1
- Application number: 18/963,958
- Filed: 2024-11-29
- Published: 2026-03-05
- Earliest claimed: 2024-09-05
- Assignees: Dell Products L.P.
- Applicants: Dell Products L.P.
- Inventors: William Jeffery White; Xin Ma; George Currie
- Web page: https://patentgrep.com/patent/US20260065093A1

## CPC

- G06F 16/285: {Clustering or classification}
- G06N 3/045: Combinations of networks
- G06N 3/098: Distributed learning, e.g. federated learning
- G06N 5/022: Knowledge engineering; Knowledge acquisition
- G06N 5/04: Inference or reasoning models
- G06N 5/043: Distributed expert systems; Blackboards
- G06Q 10/087: Inventory or stock management, e.g. order filling, procurement or balancing against orders
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR

## Abstract

The disclosure leverages an endpoint management system, a local small foundation model, and a centralized foundation model to form an AI hybrid distributed inference system. The AI hybrid distributed inference system is used to provide directions based on data inputted from a system including factories, data centers, or other businesses. The system is hybrid because it leverages discriminative AI models and generative AI models. The system is distributed because the endpoint management system and the local small foundation model are located on an edge site where the data is generated, and the centralized large foundation system is located in a centralized system that is connected to a plurality of other edge sites. The endpoint management system interacts with the input data first and is used for simple issues. The local small foundation system is used for complicated issues. The centralized large foundation system is used for the most complicated issues.

## Claims

### Claim 1

A method of operating an artificial intelligence (AI) hybrid distributed inference system, comprising:
receiving first input data at an endpoint management system of the AI hybrid distributed inference system;
performing a first classification on the first input data to form first endpoint classification results;
after forming the first endpoint classification results:
making a first determination that the first endpoint classification results cannot be resolved to an existing class of a plurality of existing classes;
based on the first determination, transmitting a first query based on the first input data and the first endpoint classification results to a local small foundation system;
obtaining first retrieval augmented generation (RAG) data that relates to the first query from an enhanced distributed RAG system at the local small foundation system;
performing a first inference, by a local small foundation model of the local small foundation system, on the first RAG data to form first local classification results;
after forming the first local classification results, making a second determination that the first local classification results are resolved to any existing class of the plurality of existing classes and do not need to be sent to a centralized large foundation system;
in response to the second determination:
transmitting the first endpoint classification results and the first local classification results to a knowledge acquirer; and
performing, by the endpoint management system, a first operation in response to the first local classification results.

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
receiving second input data at the endpoint management system;
performing a second classification on the second input data to form second endpoint classification results;
after forming the second endpoint classification results, making a third determination that the second endpoint classification results can be resolved to any existing class of the plurality of existing classes;
in response to the third determination:
transmitting the second endpoint classification results to the knowledge acquirer; and
performing, by the endpoint management system, a second operation in response to the first local classification results.

### Claim 3 (depends on 1)

The method of claim 1, further comprising:
receiving second input data at the endpoint management system;
performing a second classification on the second input data to form second endpoint classification results;
after forming the second endpoint classification results:
making a third determination that the second endpoint classification results cannot be resolved to any existing class of the plurality of existing classes;
based on the third determination, transmitting a second query based on the second input data and the second endpoint classification results to the local small foundation system;
obtaining second RAG data that relates to the first query from the enhanced distributed RAG system at the local small foundation system;
performing a second inference on the second RAG data to form second local classification results;
after forming the second local classification results, making a fourth determination that the second local classification results are resolved to any existing class of the plurality of existing classes but require a centralized action;
in response to the fourth determination:
transmitting the second local classification results to the centralized large foundation system;
performing the centralized action in response to the second local classification results;
transmitting the second endpoint classification results, the second local classification results, and action data associated with the centralized action to the knowledge acquirer; and
performing, by the endpoint management system, a third operation in response to the second local classification results.

### Claim 4 (depends on 1)

The method of claim 1, further comprising:
receiving second input data at the endpoint management system;
performing a second classification on the second input data to form second endpoint classification results;
after forming the second endpoint classification results:
making a third determination that the second endpoint classification results cannot be resolved to any existing class of the plurality of existing classes;
based on the third determination, transmitting a second query based on the second input data and the second classification results to the local small foundation system;
obtaining second RAG data that relates to the second query from the enhanced distributed RAG system at the local small foundation system;
performing a second inference on the second RAG data to form second local classification results;
after forming the second local classification results, making a fourth determination that the second local classification results are not resolved to any existing class of the plurality of existing classes;
in response to the fourth determination,
transmitting a third query based on the second input data and the second local classification results to the centralized large foundation system;
performing a third inference using global RAG data based on the third query to form first central classification results;
after forming the first central classification results:
performing a centralized action in response to the first central classification results;
transmitting the second endpoint classification results, the second local classification results, the first central classification results, and action data associated with the centralized action to the knowledge acquirer; and
performing, by the endpoint management system, a second operation in response to the first central classification results.

### Claim 5 (depends on 1)

The method of claim 1, wherein the endpoint management system comprises:
a discriminative AI model configured to make the first endpoint classification results;
a sensor configured to gather the first input data; and
a processor configured to verify the first endpoint classification results and perform the first operation.

### Claim 6 (depends on 1)

The method of claim 1, wherein the endpoint management system comprises:
a heuristic model configured to make the first endpoint classification results;
a sensor configured to gather the first input data; and
a processor configured to verify the first endpoint classification results and perform the first operation.

### Claim 7 (depends on 1)

The method of claim 1, wherein the local small foundation system comprises:
an endpoint management system manager configured to receive the first query from the endpoint management system;
a RAG input manager configured to obtain the first RAG data from the enhanced distributed RAG system;
a prompt generation manager connected to the endpoint management system manager and the RAG input manager and configured to develop a prompt using the first RAG data that relates to the first query;
a local small foundation model connected to the prompt generation manager, the local small foundation model comprising generative AI and configured to perform the first inference on the RAG data to form first local classification results with the prompt as an input and the first local classification results as an output; and
an output guardrail module connected to the local small foundation model and configured to transmit data from the local small foundation system to the centralized large foundation system and the endpoint management system.

### Claim 8 (depends on 7)

The method of claim 7, wherein making the second determination comprises:
determining, by the output guardrail module, the first local classification results are not hallucinatory by identifying the first local classification result corresponds to a possible classification on a whitelist;
determining, by the output guardrail module, that the first local classification results are resolved to the existing class by identifying that the first local classification result does not correspond to a possible classification on a blacklist; and
determine that the first local classification results do not require a centralized action at the centralized large foundation system.

### Claim 9 (depends on 4)

The method of claim 4, wherein making the fourth determination comprises:
determining, by an output guardrail module of the local small foundation system, the second local classification results is not hallucinatory by identifying that the first local classification results correspond to a possible classification on a whitelist of possible classifications;
determining, by the output guardrail module, that the second local classification result is not resolved to the existing class by identifying that the second local classification result corresponds to a possible classification on a blacklist of classifications that cannot be resolved; and
determine that the second local classification result requires the centralized action at the centralized large foundation system.

### Claim 10 (depends on 9)

The method of claim 9, wherein the centralized large foundation system, comprises:
a guardrail module configured to receive and transmit data to the local small foundation system;
a global federated RAG module configured to receive data to produce Global RAG data;
a centralized large foundation model connected to the guardrail module and the global federated RAG module, the centralized large foundation model comprising generative AI and configured to make inferences with central classification results as an output with data received from the local small foundation system and the Global RAG data as inputs; and
a state tracking module connected to the guardrail module and the centralized large foundation model, the state tracking module configured to perform a centralized action.

### Claim 11 (depends on 1)

The method of claim 1, wherein the AI hybrid distributed inference system comprises:
a centralized system, comprising:
the centralized large foundation system configured to provide centralized services to a plurality of edge sites,
an external information repository connected to the edge sites to provide a centralized location to store data generated at the edge sites and the centralized system;
the plurality of edge sites comprising:
the local small foundation system;
the knowledge acquirer;
the enhanced distributed RAG system; and
the endpoint management system, wherein the plurality of edge sites are connected to the centralized system.

### Claim 12 (depends on 11)

The method of claim 11, wherein the endpoint management system is one of a plurality of endpoint management systems connected to the local small foundation system and the knowledge acquirer.

### Claim 13 (depends on 12)

The method of claim 12, wherein each endpoint management system of the plurality of endpoint management systems is located in a factory of a plurality of factories and one endpoint management system of the plurality of endpoint management systems comprises a sensor to gather the first input data from the factory and the first endpoint classification results and the first local classification results are determined to address issues in the factory.

### Claim 14 (depends on 13)

The method of claim 13, wherein issues in the factory comprise at least one of defects in manufactured goods, reduction in inventory, reduction in production rate, and repairs needed in manufacturing equipment.

### Claim 15 (depends on 10)

The method of claim 10, wherein the AI hybrid distributed inference system is located in a factory and the centralized action comprises ordering parts for inventory and tracking an amount of parts ordered in a central location.

### Claim 16 (depends on 10)

The method of claim 10, wherein the enhanced distributed RAG system is connected to a distributive data management module which allows communication and data sharing between the global federated RAG module of the centralized large foundation system and obtaining data from web sources.

### Claim 17 (depends on 11)

The method of claim 11, wherein knowledge acquirer comprises:
a RAG knowledge acquirer configured to synthesize data from the local small foundation system into data for the enhanced distributed RAG system;
a distributed learning knowledge acquirer configured to synthesize the data from the local small foundation system to train discriminative AI models for the endpoint management system; and
a digital twin system connected to the distributed learning knowledge acquirer and configured to rank the data from the local small foundation system for use in training the distributive AI models.

### Claim 18 (depends on 1)

The method of claim 1, wherein the local small foundation system has less than 10 billion parameters.

### Claim 19

A method of operating an artificial intelligence (AI) hybrid distributed inference system, comprising:
receiving a query from an endpoint management system at a local small foundation system of the AI hybrid distributed inference system, wherein:
the endpoint management system performs a first classification on input data to form endpoint classification results,
a first determination is made that the endpoint classification results cannot be resolved to any existing class of a plurality of existing classes, and
the query based on the input data and the first classification is transmitted to the local small foundation system to obtain a classification result that can be resolved to any existing class;
obtaining retrieval augmented generation (RAG) data that relates to the query from an enhanced distributed RAG system at the local small foundation system;
performing an inference on the RAG data to form local classification results;
after forming the local classification results, making a second determination that the local classification results are resolved to any existing class of the plurality of existing classes and do not need to be sent to a centralized large foundation system; and
in response to the second determination:
transmitting the endpoint classification results and the local classification results to a knowledge acquirer, wherein the endpoint management system performs an operation in response to the local classification results.

### Claim 20

A method of operating an artificial intelligence (AI) hybrid distributed inference system, comprising:
receiving a first query from an endpoint management system at a local small foundation system of the AI hybrid distributed inference system, wherein:
the endpoint management system performs a first classification on input data to form endpoint classification results,
a first determination is made that the endpoint classification results cannot be resolved to any existing class of a plurality of existing classes, and
the first query based on the input data and the first classification is transmitted to the local small foundation system to obtain a classification result that can be resolved to any existing class;
obtaining retrieval augmented generation (RAG) data that relates to the first query from an enhanced distributed RAG system at the local small foundation system;
performing a first inference on the RAG data to form local classification results;
after forming the local classification results, making a second determination that the local classification results are not resolved to any existing class of the plurality of existing classes and need to be sent to a centralized large foundation system;
in response to the second determination:
transmitting a second query based on the input data and the local classification to the centralized large foundation system, wherein:
the centralized large foundation system performs a second inference using global RAG data based on the second query to form central classification results and performs a centralized action in response to the central classification results, and
the centralized large foundation system transmits the central classification results to the local small foundation system; and
transmitting the local classification results, the central classification results, and data associated with the centralized action to a knowledge acquirer, wherein the endpoint management system performs an operation in response to the central classification results.
