Sword is building AI systems to help more people reach the care they need, whether they are calling a health service about symptoms, receiving proactive outreach as part of a population health campaign, or seeking mental health support through an ongoing conversation. Across these settings, Sword’s goal is to make care more accessible and proactive while making safety a core part of how AI is developed and deployed.
Three areas of Sword’s work show how:
- In care navigation, Sword is scaling conversational care navigation across several European countries, using NVIDIA Nemotron open models to process incoming patient requests.
- In population health, Sword’s AI care agents support proactive outreach through campaigns designed to engage patients before care is needed.
- In mental health, Sword is using NVIDIA NeMo UserSim and NVIDIA Nemotron-Personas datasets to simulate multi-turn conversations and evaluate how Sword’s agents respond.
NVIDIA provides open-source capabilities for speech processing and user simulation; Sword brings the healthcare applications, clinical expertise, and evaluation framework.
Open-source models for national healthcare hotlines
For a patient looking for help, a phone call can be the most accessible starting point. Sword is building AI-enabled triage and care navigation for national health services to support this first step: capturing why someone is calling, assessing symptoms against the service’s clinical protocols, and connecting them to the appropriate care pathway.
Sword is testing NVIDIA Nemotron speech models for the initial speech-processing stage, converting spoken requests into text that Sword’s downstream agentic systems can work with. Sword is also testing Nemotron-based LLMs within these systems.
Speech recognition gives the care-navigation workflow a starting point in the patient’s own words, but it only supports intake; the clinical protocols, routing logic, and escalation processes sit within Sword’s application.
As Sword expands across global health systems, the aim is to make that care navigation easier to access across languages and populations. Voice lets people explain what they need through a familiar channel, including those who find digital forms difficult.
AI for proactive, population health campaigns
The same channel works in the other direction: through population health campaigns where Sword reaches out proactively to the people who need care instead of waiting for them to come to it. One example is Sword’s work with Associação Nacional das Farmácias, Portugal’s national pharmacy association, to reach people who are due their seasonal flu and COVID-19 vaccines.
Another is colorectal cancer screening in the Azores, where Sword is evaluating conversational voice approaches. The regional program data shows where people are lost: 37.9% of invited residents consent to a home test, but only 30.3% take part, because about one in five of those who receive a kit never return it.
These are two different problems: hesitation at the invitation, and practical difficulty at home. A conversation can address both in ways a static reminder cannot. Sword escalates high-risk cases to clinicians and carries the interaction context forward across every inbound or outbound conversation.
Before any of this reaches a patient, Sword tests it with NVIDIA NeMo UserSim, which generates simulated users grounded in realistic population profiles for a given locale. For triage, simulated callers describe symptoms in their own words, change their story mid-call, and push back on advice, so Sword can check that the agent follows the clinical protocol and escalates high-risk cases.
For population health campaigns, simulated recipients respond to outbound calls the way real people do: some agree, some ask questions, some decline or say they are not the right person. Sword uses the results to improve both before production traffic exists.
Advancing safer mental health AI with user simulation
Mental health is another area where voice is essential to care. For someone seeking mental health support, a huge part of the interaction depends on feeling heard and understood throughout a conversation. Chatbots trained to be a “helpful assistants” tend to become overly agreeable and sycophantic, which in turn may reinforce unhealthy, even dangerous beliefs as the conversation develops.
How it behaves across the entire exchange, along several vectors of clinical appropriateness, matters as much as any single response. To make such evaluation and analysis possible at scale, Sword needs realistic user simulation that exposes models to a wider range of situations and interaction styles that better mirror real patients.
Sword’s MindEval2 research, developed with PhD-level licensed clinical psychologists, brings these challenges together. MindEval2, a follow-up to the original MindEval paper, proposes a new framework for automatic evaluation of clinical competence that rests on two components: MindJudge, for evaluation, and MindSim, for patient simulation.
For MindJudge, Sword develops a set of Mental Health Quality Rubrics (MHQR), which break down response quality into explicit operational criteria that are more interpretable and actionable than existing evaluation taxonomies. MindSim is an LLM harness based on NVIDIA NeMo UserSim that combines carefully designed patient profiles (drawing on NVIDIA Nemotron-Personas) with tools that enable more varied and challenging interactions that are more realistic than existing baselines.
The harness contains a memory component that is updated between each interaction, enabling MindEval2 to model the longitudinal aspect of mental health. Sword evaluates 18 proprietary and open-weight models in MindEval2, finding that while frontier models have largely improved on issues related to contextual fidelity and boundaries, guidance calibration remains the dominant failure mode across all models, accounting for most of the quality gap.
Components of this framework are being used not only for evaluation of models, but are also central to a range of other applications: identifying behaviors that need improvement, large-scale synthesis of environments for reinforcement learning (RL), and even RL itself as Sword scales its use of multi-turn environments that require synthetic users. Going further, Sword is also exploring Nemotron 3.5 Super VL to act as the synthetic patient in these very same use cases.
Building safety into the path to scale
Patient triage, population health screening, and mental health dialogue serve different needs, but all three depend on AI that can handle real conversations responsibly. NVIDIA Nemotron open models support how Sword receives patient requests, while simulation tools help it rigorously examine and improve how conversational systems behave before wider use.
By combining these capabilities with clinical expertise, Sword is building toward more accessible patient-facing agentic systems, with clearer evidence to guide their improvement. The goal is to help more people reach appropriate care while making clinical oversight and rigorous evaluation part of how those systems scale.
