-
-
main pic
-
targeted retrieval demo
Project overview
This project develops a conversational AI assistant that supports people who experience loneliness, isolation, or limited access to mental health resources. Survey data reports loneliness among 79% of people in the current generation. The number of close friendships declined between 1990 and 2021, while suicide rates have increased since the 2000s.
The assistant provides emotional support and reflection. It does not replace professional care or eliminate loneliness. Its purpose is to reduce strain and support communication over time.
Design
The system addresses four limitations of standard conversational AI: limited long-term memory, weak short-term relevance, generic chatbot language, and the absence of proactive follow-up.
After each conversation, specialized GPT agents update summaries across categories, including well-being, relationships, health, routines, goals, challenges, communication style, and interests. A master summary combines the updated information for future conversations.
During each interaction, a retrieval agent selects up to three relevant summaries. Another agent integrates that information so responses reflect the user’s context.
The assistant uses three layers to improve communication. A master prompt defines a conversational SMS style. The system reads the user’s preferred communication style and updates that preference after each interaction. A humanization filter changes responses based on the previous ten messages, separates long responses, and removes unnecessary punctuation.
A scheduling agent evaluates whether a follow-up message could help. If appropriate, it creates and schedules a check-in within one to 24 hours.
Development and results
The project required 35 prompts across 10 specialized agents. Key challenges included maintaining consistent formats for agent output, coordinating agent workflows, and producing a natural conversational style. The system supports targeted memory retrieval, coordinated summaries, personalized responses, and scheduled check-ins.
Future work includes improving retrieval and scheduling, sending relevant memes, and adding scheduled prompts that refine how the assistant communicates with each user.
