AI companion architecture for elderly patients

The challenge

HumanIA Care's mission is to bridge the digital divide for elderly patients. Their product team had a clear vision: a conversational companion that could answer health questions, help with everyday tasks, and adapt to each user's vocabulary and pace. The system had to be reliable in low-bandwidth conditions, handle noisy speech input, and never confuse users who had little tolerance for technology errors.

Healthcare context added a further constraint: every response had to be grounded, auditable, and consistent with clinical guidelines. Hallucinated content was not an acceptable failure mode. Sovereign deployment — with zero cloud dependency — was a non-negotiable requirement from the start.

What we built

We designed a multi-agent AI architecture with a fine-tuned speech recognition model optimised for elderly vocal patterns. On top of this we deployed a retrieval-augmented generation system using LangChain and Weaviate, grounded in curated, clinically reviewed knowledge. A personalisation engine adapted response style, length, and vocabulary for each user over time. The entire system runs as a multi-agent pipeline with RAG, voice, and memory layers orchestrated across 16 Nexus MDS Core services.

Privacy was built in from day one: all personal data anonymised at ingestion, full GDPR-compliant data pipeline, and end-to-end audit logging. The system was deployed on-premise with containerised Kubernetes workloads and zero cloud dependency, keeping response latency under 800ms at the 99th percentile even on mobile connections.

Results

Six months from first commit to production-ready architecture. The RAG architecture eliminated hallucinated clinical advice entirely, achieving zero documented incidents in testing. The platform is designed for scale in regulated care environments and is currently in controlled deployment — not yet in public rollout.

Technologies used