TECHNOLOGY & RESEARCH
The architecture we are exploring, the research informing it, and the work needed to evaluate it.
Medugentics is being designed to reconstruct relevant context as a question develops. The goal is a connected explanation that preserves the path back to its sources.
Begin with the question and the information the person chooses to share.
Find relevant passages and retain their source, date, and relationship to the question.
Follow connections, compare conflicting information, and identify gaps.
Carry useful context into the next conversation, with review and correction built into the intended experience.
These findings concern external research systems. They are not performance claims for Medugentics or evidence of readiness for routine clinical use.
Diagnosis on difficult cases. Microsoft reported 85.5% accuracy across a benchmark of 304 New England Journal of Medicine (NEJM) cases with the Microsoft AI Diagnostic Orchestrator (MAI-DxO) and OpenAI’s o3 model, versus a mean of 20% for physicians working without their usual supporting tools.
Read the original research ↗Dialogue that listens. Google’s Articulate Medical Intelligence Explorer (AMIE) research evaluated diagnostic conversations in blinded consultations with simulated patients. It reported strong diagnostic and communication results against primary-care physicians in that setting
Read the original research ↗Open models. Google reports 87.7% zero-shot accuracy on the four-option Medical Question Answering (MedQA) benchmark for MedGemma 27B text-only.
Read the original research ↗Memory that reasons. MRAgent, a research framework for reconstructing memory, reports improvements of up to 23% over strong baselines on long-memory benchmarks, with reduced token and runtime costs.
Read the original research ↗Product evaluation must examine source accuracy, relevance, handling of uncertainty, privacy controls, usability, and performance in each intended setting. Research benchmarks do not substitute for that work.
Health, research, learning, workplace knowledge, and legal research share a need to connect fragmented information with a specific question. Each would require its own sources, safeguards, and validation.