AURA - Autonomous Unified Research Agent

AURA is a multi-agent research pipeline built to attack hallucination, slow document retrieval and untrusted code execution in automated AI research. A controller agent decomposes the query into a DAG, an embedding agent generates dual vector embeddings (OpenAI text-embedding-3-small with an automatic Gemini fallback), and a hybrid retrieval agent fuses BM25 keyword scoring with Supabase pgvector HNSW cosine similarity using Reciprocal Rank Fusion and time-aware decay. A claim verification agent triangulates facts and scores claim confidence, while a sandbox execution agent runs code in a resource-capped Docker container with no network access and a non-root user. Before any agent executes sandboxed code or an external web action, an interactive permission modal requires explicit user authorisation. Results are presented in a minimalist dark-mode single-page interface with live agent badges, a streaming reasoning log, citation cards and a governance drawer.