AI Email Writer
Project Overview
This project customizes existing LLM models to adopt a specific writing style, creating an automated email writer that drafts emails in the user’s personal style. The application can also be tailored to specific behaviors, such as mimicking Shakespeare’s writing style, based on user preferences.
Challenge
The primary objective was to develop an AI agent capable of writing emails in a specific style and tone. Unlike Shakespeare’s well-documented writing style, personal writing styles are not inherently known to generic pre-trained LLMs like GPT-4 and Claude 3.5. A customized AI workflow was designed to address this challenge.
Methodology
- Data Extraction: Extract the emails and clean the data into the required format.
- Workflow Design: Develop a prototype AI agent by breaking down the main objective into smaller tasks.
- Prompt customization: Design the system, user and assistant prompts with customization to fit the defined workflow.
- LLM Fine-tuning: Fine-tune the GPT-4o LLM using the prepared email data and tailored prompts.
- Evaluation and Improvement: Validate the model performance on held-out data and refined the prompts as needed.
- Deployment: Launch the AI email writer application with cloud services.
Tools
The project was primarily dependant on the GPT-4o LLM, utilizing the OpenAI API in Python for workflow design and development. Git and GitHub were used for version control and collaboration. Testing and deployment were assisted by continuous integration/continuous deployment (CI/CD) practices. The Google Cloud Platform (GCP) was utilized for cloud deployment.
Results
The outcome was an AI email writer tailored to a specific writing style and behavior. Unlike the generic pre-trained LLMs, the customized AI email writer produced human-like text, significantly enhancing productivity in routine business and personal tasks.