Sphere announced TRAM (Tax Review and Assessment Model), an AI-native system that automates the interpretation of tax legislation across 100+ jurisdictions by ingesting statutory law, administrative guidance, and case law, then generating executable tax rules with expert review and continuous monitoring for regulatory changes.
RapidFire AI RAG is an open-source framework for systematically tuning Retrieval Augmented Generation pipelines by parallelizing experimentation across chunking, embedding, reranking, and prompt engineering parameters to improve grounding and reduce hallucinations in LLM applications.
NVIDIA AI-Q achieved first place on DeepResearch Bench I and II by implementing a multi-agent architecture (orchestrator, planner, researcher) powered by fine-tuned Nemotron 3 Super models, trained on 67k SFT trajectories from research synthesis tasks, demonstrating that open, modular agentic systems can achieve state-of-the-art performance on research agent benchmarks.
Demonstrates document poisoning attacks against RAG systems where malicious documents injected into vector databases can manipulate LLM outputs with 95% success rate on small corpora, and evaluates five defense layers including embedding anomaly detection which reduces attack success to 20% standalone.