CASE STUDY 01
AI PRODUCT
LumiReg
Can regulatory research become conversational without losing trust and provenance?
AI-native regulatory research product
- Regulatory content
- Metadata & jurisdiction
- Vector search
- Source-grounded answers
01 / THE PROBLEM
Answers are easy.
Trust is the product.
LLMs can answer questions. But regulated users need to know:
- Where did this answer come from?
- Which jurisdiction applies?
- Which requirement supports the conclusion?
- Can I trace the reasoning back to evidence?
A fluent answer is only useful when the researcher can inspect its sources and understand its scope.
02 / DISCOVERY
Start with the research workflow.
Ten interviews and usability sessions grounded the work in how people research regulation. The product question was how to make that process conversational while retaining the source material, jurisdiction and provenance needed to judge an answer.
03 / ARCHITECTURE
From regulation to traceable evidence.
- 01Regulations
- 02Heading-aware ingestion
- 03Metadata & jurisdiction
- 04Embeddings
- 05Vector retrieval
- 06LLM orchestration
- 07Source-grounded answer
- 08Traceable evidence
The source structure and metadata travel with the content. Retrieval supplies relevant context; the answer keeps a path back to the evidence.
04 / PRODUCT JUDGEMENT
A focused research product.
BUILT AROUND
Conversational research, source-grounded answers, jurisdiction and traceability.
PRODUCT BOUNDARY
Research support. A generated answer is evidence to inspect, rather than a substitute for professional judgement.
That boundary keeps the product focused on helping users find and assess relevant requirements. The strongest claim is a working research MVP, rather than autonomous regulatory decision-making.
05 / DELIVERY & OUTCOME
2 weeks.
From problem discovery to a working external-user MVP.
What it demonstrates: product discovery, deliberate scope and hands-on AI systems engineering in a trust-sensitive domain.