AI Health Insight Engine for Elderly Care
A digital health startup needed an AI backend that could surface meaningful health observations from longitudinal wellbeing data, grounded in clinical evidence and appropriate for elderly users without medical training.
The client
- A digital health startup
- Aged care, Digital health
- New Zealand
Overview
The client runs a platform that collects longitudinal wellbeing data for elderly users: mood check-ins, sleep patterns, activity levels, symptom logs, and medication tracking across a rolling window. Users were logging consistently, but neither they nor their care networks were getting actionable insight from the patterns in that data.
Clearlead designed and built a standalone AI backend that detects meaningful health patterns, retrieves supporting evidence from a curated clinical knowledge base, and generates plain-language observations grounded solely in that retrieved evidence, with multi-layer output validation before anything reaches a user.
The result is a production-ready system that gives elderly users and their care networks evidence-backed health observations, with structural guarantees that outputs remain observations and never become diagnoses or treatment recommendations.
Methods
- RAG
- LangGraph
- Agentic AI
- Safety-constrained AI
- Stateful AI pipelines
Engagement type
- Design and build
The situation
The client's platform serves elderly users and their families, accumulating detailed day-by-day health signals over time: mood, sleep quality, activity, self-reported symptoms, and medication adherence. The data was rich, but it sat largely uninterpreted. Users received no automated signal about what their patterns might mean, and care networks had no visibility into emerging trends before they became serious concerns.
The requirement for an AI-powered insight layer came with a strict constraint. Any system operating in this space must never diagnose a condition, never recommend treatment, and never overstate certainty. Outputs needed to be clinically responsible: framed in plain language appropriate for elderly users without medical training, and pitched at the level of an observation a GP might find worth discussing, rather than a conclusion only a clinician should draw.
That constraint had direct architectural implications. A standard generative model prompted to produce health observations would draw on its parametric training knowledge, opening the door to outputs that sound authoritative but are not grounded in verified sources. The solution had to eliminate that risk at the architecture level, not merely attempt to mitigate it through prompt design.
The engagement
Clearlead designed and built the insight engine as a standalone backend service, integrating with the client's existing infrastructure through a single well-defined endpoint. The core design challenge was structural: building a pipeline where grounding guarantees were architectural, not dependent on model behaviour alone.
-
Multi-domain pattern detection across the full health signal set, with each evaluator running independently against defined temporal baselines and only triggering when meaningful patterns are present. Evaluators span mobility and falls risk, mood and cognitive indicators, sleep and fatigue patterns, symptom clustering, and urgent escalation signals.
-
Retrieval-augmented generation against a curated clinical knowledge base built from authoritative sources covering aged care, falls prevention, mental wellbeing, and related health domains across Australia and New Zealand. Semantic search retrieves the most relevant evidence, with citation tracking maintained throughout the pipeline.
-
Evidence-constrained LLM generation where the model is permitted to draw only on evidence retrieved in that request. If retrieval does not return sufficient high-quality evidence, the insight is suppressed entirely rather than generated from general model knowledge. This eliminates the hallucination risk that would arise from allowing the model to draw on parametric training data.
-
Multi-layer output validation with pre- and post-generation suppression rules enforcing hedged, non-causal language and blocking any output that implies diagnosis, causality, or clinical certainty before the response is returned.
-
LangGraph orchestration with guardrails enforced at the graph level: pattern detection, evidence retrieval, generation, and safety validation are distinct pipeline nodes with defined contracts. The safety validation gate cannot be bypassed; the safety properties of the system are structural, not reliant on model compliance.
-
HIPAA-compliant AWS production deployment, with the service running as a stateless, containerised function: encryption at rest and in transit, IAM-authenticated API access, full audit logging, and no persistent health data between requests.
Outcomes
- Elderly users and their care networks receive plain-language health observations grounded in authoritative clinical evidence, with architectural guarantees that outputs remain observations and never become diagnoses or treatment recommendations.
- A production-ready, HIPAA-compliant AI backend on AWS with structural grounding guarantees: the LLM cannot draw on parametric knowledge, safety validation is a hard gate enforced at the LangGraph pipeline level, and no health data persists between requests.
- A planned extension covering medication-linked insights was designed into the architecture from the start, with the payload schema and pipeline gating already in place. The current engagement is complete; the extension is ready to activate when the client's electronic health record integration is in place.
- A reusable architectural pattern for AI in regulated health contexts: evidence retrieval as a safety mechanism rather than a retrieval optimisation, combined with structured suppression rules, produces a system with a clean and auditable safety story.
Similar applications
The evidence-constrained generation pattern used in this engagement applies wherever AI must produce useful, confident outputs in a domain where hallucination or overstatement carries real risk.
Patient-facing health information
GP practices or health portals where outputs must stay grounded in approved clinical guidance rather than general model knowledge.
Chronic disease management
Platforms that surface trends in patient-reported data over time, with appropriate framing for non-clinical audiences.
Mental health and wellbeing apps
Personalised observations requiring both clinical grounding and careful communication boundaries.
Insurance and financial assessments
AI-generated assessments that must be evidence-traceable and avoid implying conclusions only qualified advisors should reach.
Legal and compliance summarisation
AI summarisation grounded in specific documents, without extrapolation beyond what retrieved source material supports.
AI as decision support, not conclusion
Professional contexts where AI assists a qualified person rather than replaces them: outputs inform judgement but are not presented as advice in their own right, and human review remains a mandatory step.
Discuss a similar engagement
Book a free 30-minute call. We will answer questions directly and say no when something is not a fit.