Data Stack Academy
Data Build project

ActionResponder

Project Overview

Action Responder is an AI powered app that helps trademark attorneys respond to USPTO office actions (rejections) by using data and LLMs such as ChatGPT.

 

This project presented exciting challenges:

  • Automatically scrap live trademark information
  • Build a database of trademarks for context
  • Dynamically generated a series of LLM prompts
  • Fine-tune and score responses
  • Create an intuitive UI

AI Generated Response

The response below is fully automated through the app using GPT4.

Context:

This trademark was rejected by US Patent attorney due to “Mr Purl” (this trademark application) sounding similar to “Mrs Pearl” (another existing trademark)

 

Application #: 90777928

In comparing the sound of the applicant’s mark, “MR. PURL,” with the cited mark, “MRS PEARL,” it is clear that there are significant phonetic differences that distinguish the two when spoken aloud, thereby eliminating any likelihood of confusion.

Firstly, the initial components of the marks, “MR.” and “MRS,” are pronounced distinctly different. “MR.” is pronounced as “mister,” while “MRS” is pronounced as “missus.” The difference in sound between “mister” and “missus” is clear and distinct, with the former ending in a hard “-ter” sound and the latter ending in a softer “-sus” sound. This difference in pronunciation is significant and would be immediately noticeable to the average consumer.

Secondly, the latter components of the marks, “PURL” and “PEARL,” also have distinct phonetic differences. “PURL” is pronounced as “pur-l,” with a short “u” sound, while “PEARL” is pronounced as “purl,” with a long “e” sound. The difference in vowel sounds between “PURL” and “PEARL” is significant and creates a clear distinction in pronunciation. The short “u” sound in “PURL” is a sharper, more abrupt sound, while the long “e” sound in “PEARL” is a softer, more elongated sound. This difference in vowel sounds would be immediately noticeable to the average consumer when the marks are spoken aloud.

Furthermore, the overall rhythm and cadence of the marks when spoken aloud are different. “MR. PURL” has a more staccato rhythm due to the short “u” sound in “PURL,” while “MRS PEARL” has a more fluid rhythm due to the long “e” sound in “PEARL.” This difference in rhythm and cadence further distinguishes the marks when spoken aloud.

The Process

An attorney enters their trademark application number.

The app automatically scrapes relevant trademark information including any rejections posted of US Patent & Trademarks Office.

A brief summary of trademark and rejections is generated using AI.

The app searches an embedding database of relevant case law to prepare and contextualize GPT for a response.

It then makes multiple LLM passes to generate a full response using a series of robust and dynamic prompts.

The final response is tested for depth of argument, accuracy, and hallucinations.

Back-end Components

  • Dynamic Response Generation Core [gpt4]
  • Embeddings Database: relevant trademark context
  • Guardrails: prevent hallucinations
  • Dynamic Scrappers: Pulling 12x trademark features
  • Flask APIs
  • Infrastructure built on Google Cloud Platform

The Team

This project was delivered in collaboration with our close partner, Pandata. Pandata is a top-tier AI and Data Science agency.

Drew

Lead Data Engineer

Data Stack

Chloe

Data Engineer

Data Stack

Nicolas

Data Scientist

Pandata