Decision to improve
How can people ask questions in everyday language and still trust where every number and conclusion came from?
What shapes the outcome
Conversational interfaces make complex data easier to explore, but a fluent answer can still use the wrong definition, overstate what the evidence supports, or produce a calculation that cannot be repeated. The goal was to make analytics easier to use without making the language model the source of truth.
Constraints and uncertainty
- Every question must connect to approved measures, dimensions, data versions, and comparison rules.
- Questions about what happened, why it happened, what may happen, and what caused it require different kinds of evidence.
- The system must say when a request cannot be supported instead of inventing a plausible answer.
- The experience must stay interactive while every result remains traceable and reproducible.
How the system works
- Used the LLM to translate a user’s intent and explain results, but not to calculate the answer.
- Sent approved requests to deterministic descriptive, retrieval, cohort, and causal analysis engines.
- Attached the source, analytical variant, data version, evidence level, code, and a stable result hash to every output.
- Added monitoring, role-specific views, causal-design workflows, and an independent golden test set to the same interface.
What this system enables
Insight Harness lets people explore data conversationally, drill into results, monitor changes, and download evidence-backed outputs. It also refuses unsupported requests and supports registered causal designs, so convenience does not come at the expense of control.
What I built
I designed and implemented the full governed AI system: the product architecture, shared metric definitions, deterministic analysis engines, evidence contract, evaluation harness, and interactive workbench.