The End of Black-Box AI.

Hearth Insights provides the Holt platform: a system for auditable, graph-based agentic workflows designed specifically for highly regulated environments. In healthcare, finance, and bioinformatics, an AI system is only as defensible as its audit trail. We change the paradigm by ensuring every single action is recorded on an immutable ledger. This guarantees compliance, whether you are operating under clinical validation requirements or within the FCA Digital Sandbox. Current automated pipelines routinely fail enterprise implementation because they cannot prove how they arrived at an answer. Holt enforces absolute provenance.

How does Holt provide an audit trail for clinical evaluation?

Our architecture is not theoretical. In a recent peer-reviewed study published in the Gut appendix, Holt was tested in UHSFT to detect HPO terms from gastrointestinal clinical notes using an agentic workflow.

• The Challenge: Manual coding is too slow, but existing automated pipelines lack the comprehensive audit trails required for regulated clinical environments.
• The Holt Engine: A workflow utilising four agents operating in a secure, air-gapped environment with minimal privileges. Every step: processing text, retrieving information, creating prompts, and executing the model, was recorded immutably.

Why do compact edge models outperform massive, resource-heavy models in graph-based workflows?

The study yielded a critical architectural insight: when deployed within an auditable, graph-based pipeline, massive parameter counts are not a prerequisite for high accuracy in complex clinical extraction tasks.

When tested across 15 different large language models, a compact 3-billion parameter edge model (Ministral-3) outperformed a massive 120-billion parameter model (GPT OSS) in overall precision and recall. Later testing confirmed that while model architecture remains important—with models like Medgemma performing exceptionally well—Holt’s structured workflows allow smaller, targeted models to achieve enterprise-grade results.

By orchestrating AI through Holt and decomposing complex clinical workflows into discrete, verifiable steps, institutions can deploy highly secure, locally-hosted edge models that deliver superior accuracy. Furthermore, executing this workflow through Holt proved significantly faster than standard Python-only implementations. The result is a substantially reduced AI carbon footprint, eliminated compute bottlenecks, and zero data egress.