DeepKeep Multi-Lingual Guardrails for Safe AI-Based Health Data Analytics in Low-Value Care (CARE-03)

BS"D

DeepKeep is an EIC Accelerator-winning startup providing algorithmic security and trustworthiness technology for AI systems at development and deployment stages.

For HORIZON-HLTH-2026-01-CARE-03: Identifying and addressing low-value care in health and care systems, proposals may include digital and artificial intelligence-based analysis, models and/or tools to identify and address low-value care. DeepKeep offers a practical, low-friction contribution: integrating our existing multilingual, context-aware, guardrails/filter layer around AI components used in the project—particularly where outputs can influence decisions, recommendations, or prioritisation.

What we offer to contribute:

• Multi-Lingual PII leakage prevention for AI/LLM-assisted interfaces that query, summarise, or explain results (e.g., analyst/clinician-facing assistants).

• Safety and policy filtering to reduce harmful or inappropriate outputs and enforce project-defined constraints (e.g., “no patient-identifying content,” “no unsupported clinical advice”).

• Integration support: deploying the guardrails as an add-on layer to the consortium’s chosen AI components (APIs / gateways / wrappers, depending on the stack).

• Evaluation support: helping the consortium define and run tests that demonstrate reduction in leakage and unsafe outputs (evidence useful for adoption and governance).

What we are looking for:
Consortia using AI/LLM components as part of data models, digital tools, or AI-based analysis related to low-value care, who want an implementation-ready trustworthiness layer with multi-lingual