Inspiration
With AI agent systems on the rise, their trust and safety is a concern with over 68% of users reporting they are unsure whether agent-generated outputs can be trusted without oversight and governance.
What it does
It automatically generates research ideas grounded in real papers while ensuring each idea is actually novel and safe. It combines NOVA with AEGIS that reviews, attacks, and improves the outputs. It produces full traces and safety checks so users get novel research ideas.
How we built it
We designed NOVA to retrieve and score academic papers and synthesise new research ideas. We built AEGIS as a dual-agent oversight layer with a Guardian and Attacker to validate, stress-test, and refine NOVA’s outputs. We integrated a feedback loop where AEGIS continuously improves NOVA, creating a self-improving meta-pipeline.
Challenges we ran into
Ensuring that ideas were genuinely original and not accidentally plagiarised from existing literature. Getting the Attacker to produce meaningful adversarial prompts without destabilising the entire system. Designing a governance layer that is strict enough to improve NOVA but flexible enough to keep creativity alive.
Accomplishments that we're proud of
We built a working system where two agents collaborate and challenge each other to produce safer, smarter research ideas. We managed to reduce hallucinated citations, catch flawed reasoning, and increase novelty through automated oversight.
What we learned
Creativity alone isn’t enough governance is essential for making AI reliable, especially in research settings. Adversarial testing dramatically improves idea quality and exposes failure modes single agents can't catch. Building trust in AI systems requires visibility into not just the output, but the entire reasoning journey.
What's next for Prism - X
Expanding the system to handle multimodal inputs like datasets, graphs, or code, not just academic text. We have also stated further development plans in our research paper
Log in or sign up for Devpost to join the conversation.