# US 2025/0094150 A1 · Ai model for synthesizing software from a user prompt

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

- Number: US 2025/0094150 A1
- Document ID: US20250094150A1
- Status: Published application
- Kind: A1
- Application number: 18/889,716
- Filed: 2024-09-19
- Published: 2025-03-20
- Earliest claimed: 2023-09-19
- Applicants: Durable AI
- Inventors: Michael Kasper; Calvin Smith; Fernando Nobre; Nima Keivan; Theodore Hoff
- Web page: https://patentgrep.com/patent/US20250094150A1

## CPC

- G06F 8/60: Software deployment
- G06F 8/10: Requirements analysis; Specification techniques
- G06F 8/30: Creation or generation of source code
- G06F 8/35: model driven
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

Systems and methods for synthesizing software application code for a software application. An example method includes receiving, from a user device, an initial prompt comprising a natural language description of a plurality of specification elements defining a functionality and a scope of a software application; synthesizing software application code for the software application by processing the initial prompt through a synthesis model, wherein the synthesis model is trained to: generate, based on the initial prompt, a specification document comprising the specification elements; and generate the software application code for the software application by determining a bidirectional mapping between the software application code and the specification elements; and deploying, to a deployment platform, the generated software application code for execution.

## Claims

### Claim 1

A method for synthesizing software application code for a software application, the method comprising:
receiving, from a user device, an initial prompt comprising a natural language description of a plurality of specification elements defining a functionality and a scope of a software application;
synthesizing software application code for the software application by processing the initial prompt through a synthesis model, wherein the synthesis model is trained to:
generate, based on the initial prompt, a specification document comprising the plurality of specification elements; and
generate the software application code for the software application by determining a bidirectional mapping between the software application code and the plurality of specification elements; and
deploying, to a deployment platform, the software application code for execution.

### Claim 2 (depends on 1)

The method of claim 1, wherein the synthesis model is trained to generate the specification document by:
providing, to the user device, a prompt for clarification generated by processing the initial prompt through the synthesis model; and
updating the specification document by processing an answer to the prompt for clarification, received from the user device, through the synthesis model.

### Claim 3 (depends on 2)

The method of claim 2, wherein the answer to the prompt for clarification includes a missing fact regarding the software application.

### Claim 4 (depends on 2)

The method of claim 2, wherein the answer to the prompt for clarification includes an approval or disapproval of a statement of assumption.

### Claim 5 (depends on 2)

The method of claim 2, wherein the prompt for clarification includes a statement of an assumption provided by the synthesis model.

### Claim 6 (depends on 2)

The method of claim 2, wherein the prompt for clarification includes a conflicting requirement, a contradictory requirement, or an impossible requirement.

### Claim 7 (depends on 6)

The method of claim 6, wherein the answer to the prompt for clarification includes a resolution for the conflicting requirement, the contradictory requirement, or the impossible requirement.

### Claim 8 (depends on 1)

The method of claim 1, wherein the synthesis model is trained using a supervised learning loss function that rewards a correct use of knowledge base entries, stored to a knowledge base, given a context of a synthesis problem and the plurality of specification elements being considered.

### Claim 9 (depends on 8)

The method of claim 8, further comprising:
storing the initial prompt and an answer to a prompt for clarification to the knowledge base as knowledge base entries.

### Claim 10 (depends on 1)

The method of claim 1, further comprising:
providing, to the user device, a plurality of validation scenarios to assess a correctness of the software application code;
receiving, from the user device, one or more rejections of one or more of the plurality of validation scenarios; and
providing, to the user device, a prompt to modify the specification document based on the one or more rejections of one or more of the plurality of validation scenarios.

### Claim 11 (depends on 1)

The method of claim 1, further comprising:
receiving, from the user device, a modification to the initial prompt or an answer to a prompt for clarification within the specification document.

### Claim 12 (depends on 11)

The method of claim 11, wherein the synthesis model is trained to update the software application code based on the modification or alternatively emit further clarifying questions, statements of assumptions, or issues.

### Claim 13 (depends on 1)

The method of claim 1, further comprising:
receiving, from the deployment platform, a runtime error generated during the execution of the software application code;
determining, based on the bidirectional mapping, a modification to the specification document to correct the runtime error; and
providing a prompt for the modification to the user device.

### Claim 14 (depends on 1)

The method of claim 1, wherein the software application comprises frontend components and backend components.

### Claim 15 (depends on 1)

The method of claim 1, wherein the software application code is configured to, when executed by the deployment platform, perform at least one action, and then terminate.

### Claim 16 (depends on 1)

The method of claim 1, wherein the software application code is configured to, when executed by the deployment platform, continuously perform at least one action.

### Claim 17 (depends on 1)

The method of claim 1, wherein the synthesis model comprises a retriever model and a generator model that are trained with synthetic data spanning domains of natural language with an internal representation.

### Claim 18 (depends on 1)

The method of claim 1, further comprising:
retraining the synthesis model with the initial prompt, the specification document, or the software application code.

### Claim 19 (depends on 1)

The method of claim 1, wherein the synthesis model comprises a trained artificial intelligence (AI) model.

### Claim 20 (depends on 19)

The method of claim 19, wherein the synthesis model comprises a trained neural network.

### Claim 21 (depends on 1)

The method of claim 1, wherein the plurality of specification elements are provided in natural language in the specification document.

### Claim 22 (depends on 1)

The method of claim 1, further comprising:
receiving, from the user device, a request to supervise an execution of a step related to at least one of the plurality of specification elements, wherein the step is determined according to the bidirectional mapping;
before deploying the software application code to the deployment platform, integrating the request into the software application code;
providing, to the user device, a prompt for supervising of the step received from the deployment platform during the execution of the software application code; and
providing, to the deployment platform, a supervision instruction related to the step received from the user device in response to the prompt for approving or rejecting the step.

### Claim 23 (depends on 22)

The method of claim 22, further comprising:
providing, to the user device, a prompt to modify the specification document when the supervision instruction includes rejecting of the step.

### Claim 24

A system for synthesizing software application code for a software application, the system comprising:
a user device;
a deployment platform; and
an electronic processor configured to:
receive, from the user device, an initial prompt comprising a natural language description of a plurality of specification elements defining a functionality and a scope of a software application;
synthesize software application code for the software application by processing the initial prompt through a synthesis model, wherein the synthesis model is trained to:
generate, based on the initial prompt, a specification document comprising the plurality of specification elements; and
generate the software application code for the software application by determining a bidirectional mapping between the software application code and the plurality of specification elements; and
deploying, to the deployment platform, the software application code for execution.

### Claim 25

A non-transitory computer readable medium having stored thereon executable instructions that, when executed by an electronic processor, cause the electronic processor to perform operations comprising:
receiving, from a user device, an initial prompt comprising a natural language description of a plurality of specification elements defining a functionality and a scope of a software application;
synthesizing software application code for the software application by processing the initial prompt through a synthesis model, wherein the synthesis model is trained to:
generate, based on the initial prompt, a specification document comprising the plurality of specification elements; and
generate the software application code for the software application by determining a bidirectional mapping between the software application code and the plurality of specification elements; and
deploying, to a deployment platform, the software application code for execution.
