Ouroboros

Ouroboros is a calm, context-aware AI guide that explores ideas with users, using knowledge-driven conversations while respecting silence and avoiding off-topic chatter.

Comment

Inspiration

Many conversational AIs either repeat themselves, stray off-topic, or insist on responses, making conversations frustrating. I wanted to build an AI that respects context, encourages reflection, and guides thoughtful exploration without overwhelming the user.

Ouroboros was inspired by the idea of a calm, persistent guide that learns from interactions while staying grounded in verified knowledge. Instead of treating AI as a machine that must always answer, Ouroboros explores the possibility of human-centered conversational AI that respects silence, topic shifts, and curiosity.

What it does

Ouroboros is a context-aware conversational AI guide that helps users explore ideas thoughtfully while staying grounded in knowledge.

It uses Amazon Nova 2 foundation models through Amazon Bedrock combined with a retrieval-augmented generation (RAG) approach. Before generating a response, Ouroboros retrieves relevant information from a knowledge base, ensuring answers remain accurate and focused.

Key capabilities include:

  • Using a private knowledge base to provide accurate, grounded responses
  • Tracking conversation history to avoid repeating questions
  • Respecting user silence or topic changes naturally
  • Avoiding straying into off-topic or unknown subjects
  • Maintaining a calm and reflective conversational tone

How I built it

Ouroboros combines foundation models with structured conversation logic and knowledge retrieval.

  • AI Model: Amazon Nova 2 via Amazon Bedrock powers reasoning and dialogue generation
  • Architecture: Retrieval-Augmented Generation (RAG) to ground responses in knowledge
  • Backend: FastAPI application managing sessions, history, and user state
  • Cloud Infrastructure: Hosted on Amazon EC2
  • Memory tracking: Python dictionaries track which topics have been discussed
  • Knowledge retrieval: Bedrock Knowledge Base integration retrieves relevant information before response generation
  • Instruction tuning: Custom system prompts enforce natural, context-aware conversation and prevent repetition

Example logic rule used in the system:

[ \text{If user_history = \emptyset, respond with “No prior history” and store it as valid context.} ]

This combination allows Ouroboros to remain knowledge-grounded while maintaining conversational continuity.

Challenges I ran into

  • Balancing openness and restriction: Ensuring Ouroboros could explore ideas without hallucinating or inventing knowledge
  • Memory edge cases: Handling empty histories, repeated topics, and session interruptions
  • Instruction tuning: Translating human conversational norms into system prompts that the model reliably follows
  • Conversation control: Preventing the AI from over-talking while still maintaining helpful dialogue

Accomplishments that I'm proud of

  • Built a conversational AI that feels calm, reflective, and natural
  • Successfully implemented context tracking to prevent repeated questions
  • Integrated Amazon Nova 2 with a knowledge-grounded RAG architecture
  • Ensured the AI stays focused and avoids hallucinated information
  • Created a conversational system that respects user interaction patterns, including silence and topic shifts

What I learned

Building Ouroboros reinforced several important lessons about AI systems:

  • I learned how to use a suite of Amazon tools to build my first AI application, including Amazon Bedrock, Nova foundation models, and EC2 cloud infrastructure
  • Precise instructions often shape AI behavior more than code
  • Memory and context management are essential for coherent dialogue
  • Grounding models with retrieval systems reduces hallucinations
  • Small UX improvements—like acknowledging silence or topic changes—can significantly improve the user experience

What's next for Ouroboros

Next steps focus on expanding the system's capabilities while maintaining thoughtful conversation design:

  • Expand the knowledge base for broader and deeper domain coverage
  • Improve longer-term context retention and multi-turn reasoning
  • Introduce user personalization and adaptive conversational styles
  • Explore integrations with additional AI tools to support richer knowledge exploration
  • Experiment with structured memory systems to make conversations more context-aware over time

P/S: I lost more than three weeks getting stuck with AWS permissions, metadata retrieval logic and how to create a Nova automated greeting on load without passing user as “user” — wish I had more time tweaking instructions. P/P/S: The website isn’t mobile-friendly at the moment.

Built With

Share this project:

Updates