The challenge
A managed service provider assisting NDIS disability care providers required enterprise-grade specifications to assess replicating Lumary, Australia's leading NDIS care management platform. Lumary operates on Salesforce with five integrated modules, 200,000+ users, and government API integration. It processes over $6 billion in annual disability funding.
Scope included five product modules, NDIA government API integration, Australian Privacy Act and NDIS Practice Standards compliance, 99.99% uptime SLAs, and a claiming engine requiring 99%+ accuracy, where inaccuracy prevents provider payment.
The traditional approach would take two to four weeks of senior solutions architect time at $30,000 to $60,000. The client needed delivery within days.
The approach: AI as solutions architect
Rather than treating AI as a writing tool, the team positioned the model as a collaborative solutions architect, conducting research, making architectural decisions, identifying risks, and producing deliverables through structured, iterative sessions.
01 · Research and discovery
Fifteen-plus sources were synthesised in minutes, Lumary product documentation, NDIA API specifications, Privacy Act compliance requirements, and competitive landscape analysis, cross-referencing marketing claims against government documentation to establish an objective technical foundation.
02 · Platform specification
A 22-section technical specification covered the complete platform: executive summary, five module specifications with 100+ features, system architecture, AI integration strategy, data model, integrations, security, infrastructure, non-functional requirements, risk register, and a 27-month implementation roadmap.
03 · Deep-dive on the hardest problem
The NDIS claiming engine was identified as the most complex component, involving six converging technical challenges, and a separate 20-section deep-dive specification was produced with complete solution architecture, a 14-week implementation plan, and a novel AI redundancy pattern.
The innovation: AI-driven quality redundancy
The most significant output transcended specification alone. It introduced an architectural pattern for ensuring claim accuracy in systems where errors have direct financial consequences. Three independent AI-augmented layers validate each other before financial stakes arise.
Layer 1 · AI price guide ingestion
AI parses NDIS Price Guide PDFs into structured rule configurations with confidence scores. Rules below 0.9 confidence are flagged for human review. Manual encoding that traditionally took weeks now completes in hours.
Layer 2 · Configurable rules engine
Claims traverse a 12-stage validation pipeline. Rules load from configuration rather than hard-coded implementation, so Price Guide updates become configuration deployments rather than code releases.
Layer 3 · AI cross-validator
Random claim samples receive independent assessment by an LLM against raw NDIS documentation, creating a second opinion on every sampled claim before government submission. When all three layers align, confidence reaches extremely high levels. Disagreements signal potential errors caught before financial consequences, and become training signals that improve accuracy progressively. The target is 99%+ claim accuracy validated against real provider data before any live claim submission.
Claude didn't just describe the problem, it designed the defence-in-depth architecture to solve it, including the decision matrix for handling disagreements between layers and the provider validation loop to prove accuracy before go-live.
Key takeaways
- Architect, not just writer. Maximum value emerged from directing the model to make decisions, identify risks, and design systems. It independently identified the claiming engine as the hardest problem and designed the triple-redundancy architecture without explicit prompting.
- AI designing AI quality systems. The quality-redundancy pattern, three independent layers cross-validating before submission, extends to any system prioritising accuracy with frequently-changing rules.
- Specification quality drives velocity. Detailed specifications function as foundations rather than overhead, particularly when enabling AI-assisted code generation. Each specification hour recovers multiple implementation hours.
- Domain expertise remains essential. Domain experts stay critical at key checkpoints: Price Guide rule approval, provider data verification, and go/no-go decisions. AI augments rather than replaces human expertise.