Knowledge-Net AI for Assistive-Product Matching
Azure-based reasoning with Editorial gating, audit trails, and expert training for global health
The Challenge
The client needed a global, web-based machine-reasoning system to support a critical task: matching patients with appropriate assistive products from a catalog of mobility, hearing, vision, and other aids. The matching problem is genuinely difficult — assistive products vary widely by patient age, geographic context, socio-economic situation, anatomical fit, and clinical condition; a misfit recommendation has direct consequences for the patient’s mobility, independence, and quality of life.
The client needed a system that could ingest clinical and contextual information about a patient, draw on a continuously evolving knowledge base of expert-curated product profiles, generate ranked recommendations with explicit reasoning, and crucially — allow expert clinicians around the world to collaboratively train and refine the underlying engine over time. Conventional decision-tree approaches were inadequate; the matching logic involved too many variables, too much expert nuance, and too much continuous evolution to encode as static rules. The system also needed to operate globally, with appropriate latency, security, and data-handling controls suitable for an international health body.
The Solution
Applied Decision Technologies architected a probabilistic Knowledge-Net reasoning system on Microsoft Azure — the same fundamental framework subsequently applied to the firm’s parallel national-health-systems engagement on grievance redressal but tuned for the assistive-products matching problem and operating with even tighter Editorial gating.
The data tier uses Azure Cosmos DB, chosen specifically for its native graph capabilities — Knowledge Links, Knowledge Chains, and Knowledge Nets are first-class data structures in the design and benefit directly from the underlying graph storage. The reasoning engine traverses these structures using forward and backward chaining inference methods to generate ranked product recommendations from the pool of available assistive products.
Azure Cognitive Services power the natural-language interfaces: Language Understanding Intelligence Service, Entity Linking, QnA Maker, and Recommendations API are leveraged for query interpretation and recommendation refinement. Authentication and access control are unified through Azure Active Directory’s Identity and Access Management — providing centralized identity management with the agility benefits of cloud-native security configuration.
The Knowledge Training component includes a Profile Generator (for creating new product or patient-class profiles), a Profile Tuner (for interactive expert-driven scenario testing), an Inference Visualizer (a graphical tool for visualizing match results and tuning parameters), and a Collaboration Facilitator (video, audio, and chat among geographically distributed experts during training sessions, with full session recording for review).
Critically, the design adds a stronger Editorial approval workflow than the parallel national-health variant: the Profile Tuner persists tuned changes to a temporary store, where they remain quarantined until the Editorial team explicitly approves them. Audit trails capture who made each change, when, and why; dashboards alert the Editorial team to any unapproved pending changes. The reasoning engine never operates on profile data that has not passed through human approval.
Outcomes
The engagement validated the firm’s Knowledge-Net AI framework as a reusable, domain-agnostic pattern. The same architectural primitives — Knowledge Items, Knowledge Linking, Knowledge Chaining, Knowledge Repository, Reasoning Engine, Knowledge Assimilator — proven on a national-health grievance-redressal use case were redeployed for an entirely different problem (assistive-products selection) with only domain-specific tuning. The Editorial-gate pattern in particular has become a default ADecTec design element for any AI/ML workload where human accountability is non-negotiable.
The international scope of the engagement also forced the firm to think rigorously about latency, data residency, multi-language support, and the trade-offs between centralized and federated architecture — disciplines that have informed every subsequent international and government-tier ADecTec engagement.
Why It Matters
Two takeaways. First, ADecTec’s Knowledge-Net framework is genuinely portable across consequential matching problems — public-sector grievance routing, global assistive-products selection, criminal-justice defendant matching, and beyond — because the underlying architectural primitives are domain-agnostic. Second, the rigor that an international health body brings to Editorial approval, audit-trail discipline, and human-in-the-loop oversight is now baked into the firm’s default delivery posture for any AI/ML engagement.
Clients evaluating an AI/ML partner for a high-consequence matching, recommendation, or classification workload — particularly in healthcare, government, or regulated industries — get the benefit of patterns proven on engagements at international-health-body scale and rigor.