AI judge bias: why judges favor their own outputs

Rubric-based evaluation was supposed to make LLM judges objective. We show that judges keep favoring their own outputs even when the rubric is fully objective (e.g., a unit test), and we quantify what that costs when ranking frontier models.

September 23, 2026 • 12 min read

How We Cut Enterprise AI Costs by 94% Without Slowing Down Engineers

Learn how Sword Health designed a multi-agent AI system to handle member support tickets at scale — combining RAG-powered knowledge retrieval, dynamic tool orchestration, and continuous evaluation to resolve cases faster without sacrificing quality.

August 4, 2026 • 21 min read

Arbor: a Framework for reliable navigation of critical conversation flows

Navigating complex decision trees while maintaining natural conversation remains an unsolved challenge, here's how Arbor solves it.

July 28, 2026 • 10 min read

MindGuard: open-source guardrails that understand clinical context

MindGuard is a family of open-source safety classifiers for mental health conversations that reduce false positives while accurately identifying real crisis signals.

June 15, 2026 • 12 min read

How Sword built an AI care coordinator agent to scale member support

Learn how Sword Health designed a multi-agent AI system to handle member support tickets at scale — combining RAG-powered knowledge retrieval, dynamic tool orchestration, and continuous evaluation to resolve cases faster without sacrificing quality.

May 18, 2026 • 10 min read

Inside MindEval: building a benchmark for LLMs in multi-turn mental health support

How we designed, validated, and open-sourced a benchmark for evaluating LLMs in realistic multi-turn mental health conversations and what we found across 12 frontier models.

May 5, 2026 • 12 min read

How we built Phoenixʼs real-time voice experience

How we built Phoenix’s real-time voice experience, combining voice AI, clinical memory, and safety guardrails to guide members through treatment at scale.

March 13, 2026 • 12 min read