Home Case Studies Lumary NDIS Platform

Reverse-Engineering a $6B NDIS Platform in One Session

Full case study available as PDF. Includes complete specification methodology, architecture diagrams, and implementation plan.

45+
Pages of enterprise spec
1
Working session
90%
Time saved vs traditional
95%+
Cost reduction

The Challenge

A managed service provider working with NDIS disability care providers needed enterprise-grade specifications to evaluate the feasibility of replicating Lumary — Australia's leading NDIS care management platform. Lumary is a Salesforce-based product with five integrated modules, 200,000+ users, and deep integration with Australian government APIs. It processes over $6 billion in annual disability funding.

The scope was significant: five product modules, integration with NDIA government APIs, compliance with the Australian Privacy Act and NDIS Practice Standards, enterprise SLAs of 99.99% uptime, and a claiming engine that must achieve 99%+ accuracy — where errors mean providers don't get paid.

Traditional approach: 2–4 weeks of senior solutions architect time at $30,000–$60,000. The client needed it in days.

The Approach: Claude as Solutions Architect

Rather than treating AI as a writing tool, the team used Claude Opus 4 as a collaborative solutions architect — conducting research, making architectural decisions, identifying risks, and producing deliverables in a structured, iterative session.

The Innovation: AI-Driven Quality Redundancy

The most significant output wasn't just the specification — it was an architectural pattern Claude designed for ensuring claim accuracy in a system where errors have direct financial consequences. Three independent AI-augmented layers validate each other before any real money is at stake.

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. What traditionally took weeks of manual encoding now completes in hours.

Layer 2

Configurable Rules Engine

Claims pass through a 12-stage validation pipeline. Rules are loaded from configuration rather than hard-coded — so Price Guide updates are a config deployment, not a code release.

Layer 3

AI Cross-Validator

A random sample of claims is independently assessed by an LLM against raw NDIS documentation, creating a second opinion on every sampled claim before government submission.

When all three layers agree, confidence is extremely high. When they disagree, a potential error has been caught before it costs money. Disagreements become a training signal — improving system accuracy over time. Target: 99%+ claim accuracy validated against real provider data before a single live claim is submitted.

"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."

Results

Dimension Traditional AI-Driven Improvement
Research & Discovery 5–7 days Minutes ~95% faster
Platform Specification 10–15 days Single session ~90% faster
Deep-Dive Component Spec 5–7 days Same session ~90% faster
Document Formatting 2–3 days Automated ~95% faster
Estimated Cost $30,000–$60,000 Claude API + 1 session >95% reduction

Key Takeaways

Claude as architect, not just writer. The highest value came from treating Claude as a collaborative solutions architect — asking it to make decisions, identify risks, and design systems. It identified the claiming engine as the hardest problem and designed the triple-redundancy solution without being prompted to.

AI designing AI quality systems. The most innovative output was the AI-driven quality redundancy pattern — three independent layers that cross-validate before submission. A pattern applicable far beyond NDIS claiming to any system where accuracy is critical and rules change frequently.

Specification quality drives development velocity. A detailed spec isn't overhead — it's the foundation for AI-assisted code generation. Every hour invested in specification saves multiple hours in implementation.

Domain expertise remains essential. Claude accelerated the work by 90%+, but domain experts are required at critical checkpoints — Price Guide rule approval, provider data verification, and go/no-go decisions. AI augments human expertise; it doesn't replace it.

Want to see what AI-driven specification can do for your project?

Singularity Tech delivers SaaS replacements in 90 days with a written parity guarantee. The assessment is free.

Talk to our team

Download the Case Study

Enter your details to download. We'll send occasional insights — no spam, unsubscribe anytime.

Your details are only used to send the download notification.

Your PDF is downloading. Want to talk through what this means for your business?