Most knowledge workers and students spend significant amounts of their time on digital devices, on average 5.6 hours each day in Switzerland in 2023, with an increasing trend. Despite more and more tools (e.g. PersonalAnalytics, RescueTime, WakaTime, rize.io) offering to capture data on how time is spent on digital devices, providing personalized and actionable insights remains a challenge.

As a result, workers and students have only limited insights into how various aspects are impacting  their productivity and well-being. To overcome the challenge, we propose to build on top of our PersonalAnalytics platform, which already allows to capture relevant data, and allow users to write questions to query their own data. An LLM transforms these questions into data queries and analysis code that is then run on the data, and outputs insights in natural language or as simple visualizations.

That way, users can get easy and personalized insights on how well they are spending their time, on their work rhythms, or on more specific questions about particular files, websites or work habits.

Project Goals

The goal of the WorkGPT project is to build up on the PersonalAnalytics platform to …

• … leverage existing data collection from PersonalAnalytics (includes data on app usage, activity categories, time management, self-reports, and user inputs);
• … allow users to ask questions in natural language, enabling them to query their own data and receive responses in textual or graphical form; and
• … create SQL queries from user’s natural language inputs, processes the data, and outputs relevant insights using a (local or IFI-hosted) LLM.

Project Phases

• Requirement Analysis:
  • What are relevant questions that users might ask based on the available data?
  • Technical feasibility to use LLM for creating queries and analysis code
  • Technical feasibility to run LLM locally (e.g. macOS/Windows native?) to limit privacy concerns
• Implementation (main part)
  • Data Integration
  • Natural language Query Processing
  • Insight Generation and Visualization
• Preliminary evaluation + dogfooding

Note that the exact scope, features and implementation can all be discussed. This description is merely as a means to initiate discussion. Please also consider our Guidance on finding Project topics under our supervision.

Contact Andre Meyer in case of interest or questions.