Granted Patent B2

US 12,574,340 · App. 18/534,809 · Granted Mar 10, 2026

Inventors: Charles T. Natoli (San Francisco, CA); Devin Patel (Boston, MA); Sofiane Hadji (Boston, MA); Gal Gila Korcia (Cambridge, MA); Nick Vessella (Rockland, MA); Robert Huselid (Stamford, CT); Andrew Piliero (Redwood City, CA); Tristan Mills (Cambridge, MA); Harsh Mehta (Boston, MA)

Assignee: Klaviyo, Inc.

H04L51/02G06Q30/0201H04L51/063H04L51/212

Abstract

Claims (57)

1 . A computer-implemented method for generating text for short messages for electronic marketing messaging, comprising:

receiving, by a server, information related to the electronic marketing message;

receiving, by the server, a plurality of N short messages generated based on the received information from a generative text engine model;

reducing, by the server, the plurality of N short messages down to M short messages comprising:

determining similarity scores between each of the N short messages

eliminating one of any two short messages of the N short messages that are determined to have a similarity score greater than a first similarity threshold;

determining similarity scores between each of the N short messages and the received information;

eliminating short messages of the N short messages that are determined to have a similarity score greater than a second similarity threshold with the received information;

wherein the first similarity threshold is different than the second similarity threshold;

displaying messages based on the M short messages to a merchant user;

identifying merchant actions in response to the displaying of the messages based on the M short messages, comprising tracking, by the server, actions of the merchant user at a merchant server based on the messages displayed; and

fine-tuning, by the server, the generative text engine model based on the identified merchant actions, comprising:

assigning a quality rating to at least a portion of the messages based on the M short messages; and

supplementing data used to train the generative text engine model with quality ratings of the at least the portion of the messages based on the M short messages.

2 . The method of claim 1 , wherein the server is electronically connected to a merchant server of the merchant user, and electronically connected to a plurality of customer devices of customers of the merchant user, and further comprising:

displaying at least a portion of the messages based on the M short messages to one or more customers of the merchant user;

identifying customer actions in response to the displaying of the at least the portion of the messages based on the M short messages; and

further fine-tuning, by the server, the generative text engine model based on the identified customer actions.

3 . The method of claim 2 , wherein tuning the generative text engine model comprises:

assigning a second quality rating for each of the short messages displayed to the one or more customers based on the identified customer actions; and

supplementing data used to train the generative text engine model with the second quality ratings of the short messages displayed to the one or more customers.

4 . The method of claim 2 , wherein a filtering discriminator model performs at least a portion of the reducing of the N short messages down to the M short messages, and further comprising fine-tuning, by the server, the filtering discriminator model based on the identified customer actions.

5 . The method of claim 2 , wherein identifying the customer actions in response to the displaying of the at least the portion of the messages based on the M short messages comprises:

sensing the one or more customers visiting a physical location of the merchant, and the one or more customers purchasing a product or service of the merchant at the physical store location of the merchant.

6 . The method of claim 2 , wherein identifying the customer actions in response to the displaying of the at least the portion of the messages based on the M short messages comprises:

sensing combinations or sequences of customer actions; and

weighting the sensed customer actions based on the sensed customer actions; and

identifying a customer action for combinations of weighted customer actions exceeding a customer action threshold.

7 . The method of claim 2 , wherein identifying the customer actions in response to the displaying of the at least the portion of the messages based on the M short messages comprises:

identifying business locations visited by customers after receiving short messages;

rating different businesses with different customer action scores; and

identifying a customer action based on the ratings of the different businesses visited by the customers.

8 . The method of claim 2 , wherein identifying the customer actions in response to the displaying of the at least the portion of the messages based on the M short messages comprising:

tracking, by location and motion sensors of computing devices of the customers, locations and motions of the customers including tracking hand motions, direction of eyesight, and orientations of the customers;

identifying motions of a customer including how long the customer holds or looks at a specific product of the merchant;

determining a score of sensed customer actions based on sensed locations and motions of the customer; and

identifying a customer action based on the score.

9 . The method of claim 2 , wherein identifying the customer actions in response to the displaying of the at least the portion of the messages based on the M short messages comprises:

determining relationships between different customers by tracking locations of the different customers;

identifying customer friends or associates of the different customers based on location the tracking;

identifying commonalities of the different customers by identifying common locations, or common types of locations between the different customers; and

determining a success score of the M short messages based on a level of influence between the different customers.

10 . The method of claim 1 , wherein a filtering discriminator model performs at least a portion of the reducing of the plurality of N short messages down to the M short messages.

11 . The method of claim 10 , further comprising fine-tuning, by the server, the filtering discriminator model based on the identified merchant actions.

12 . The method of claim 1 , further comprising continuously updating the generative text engine model based on continuously generated quality ratings.

13 . The method of claim 1 , wherein reducing the plurality of N short messages down to M short messages comprises filtering the N short message to eliminate short messages based on content, including eliminating short messages that include inappropriate and politically sensitive words.

14 . The method of claim 1 , wherein reducing the plurality of N short messages down to M short messages comprises filtering the N short message to eliminate short messages based on content, including eliminating short messages that include identified prohibited context.

15 . The method of claim 1 , further comprising applying a fine-tuned enhancer model to enhance existing content for a selected one or more of the M short messages, wherein enhancing the content includes modifying or updating the existing content to include a branding voice of the merchant user, comprising personalizing the content for the merchant user to include the branding voice of the merchant user as determined by analyzing previous or past samples of short messages of the merchant user, and

wherein enhancing further includes adding emojis to the selected one or more of the M short messages comprising:

determining by training an enhancement model on historical short message-emoji pairs;

gathering an emoji selection by calling the enhancement model on the selected one or more of the M short messages; and

selected one or more of the M short messages according to predetermined patterns.

16 . The method of claim 15 , wherein personalizing the content for the merchant user includes personalizing the information related to the electronic marketing message before generating the N short messages, or personalizing content of the generated N short messages.

17 . The method of claim 1 , further comprising:

providing a merchant with a slider control that allows the merchant to adaptively adjust a tone of the M or N short messages comprising displaying the slider control on a user interface of the merchant allowing the merchant to adjust the tone of the M or N short messages, wherein endpoints of the slider include serious and fun.

18 . The method of claim 1 , further comprising applying a fine-tuned enhancer model to enhance existing content for a selected one or more of the M short messages, wherein enhancing the content includes modifying or updating the existing content to include a branding voice of the merchant user, comprising:

providing the merchant with a toggle that allows the merchant to selectively match a brand voice of current M or N short messages with a brand voice of past electronic messages.

Assignments (1)

Continuity (2)

References Cited (97)