When I first started coding with agents, I was deep into competitive programming and came across a provocative article titled "Building o1 with Gemini FOR FREE" (which turned out to be basic answer enhancement through agent chaining of Gemini 2.0 Flash).
Ava is a tribute to both of those formative experiences: my competitive programming background and my first encounter with AI agents. The system incorporates structured data management that agents can view and modify (like the investigation tree and case file analysis), while leveraging Gemini's free tier (still using 2.0 Flash)—staying true to that original democratized AI inspiration.
While not perfect, Ava serves as a compelling MVP of a fully automated agentic workflow for document investigation. It combines three key architectural components: an executor agent that performs document analysis, a tree-based memory system managed by a synthesizer agent that builds investigation knowledge graphs, and a task queue managed by a planner agent that coordinates analysis workflows—all working together to transform raw case files into structured investigative insights.
Log in or sign up for Devpost to join the conversation.