Conversational Analyst
Architecture

How it's built

Warehouse-agnostic by design. The NLP layer sits on top of a governed semantic model — so the same answer is correct whether your warehouse is Snowflake, Databricks, Fabric, or BigQuery.

SourcesOperational systems
  • Acumatica (ERP)
  • Ring Central
  • Senta
  • CRM
  • Workday
  • Custom APIs
WarehouseBring your own — we don't lock you in
  • Snowflake
  • Databricks
  • Fabric
  • BigQuery
Semantic + NLPGoverned metrics & natural-language reasoning
  • Metric definitions
  • Entity resolution
  • SQL generation
  • Lineage & DQ
SurfacesWhere users actually meet the answer
  • Web app
  • Slack
  • Mobile
  • Email digest
  • Power BI / Tableau

Medallion data model

Bronze for raw, Silver for cleaned, Gold for the business-facing semantic layer.

Bronze
Raw replication

CDC from every source system. Append-only, fully replayable. No business logic applied.

Silver
Cleaned & joined

Deduped, type-cast, conformed entities. Customer is one customer across Acumatica, Ring, and CRM.

Gold
Semantic / business

Governed metrics — Revenue, Margin, Service Calls — defined once, used everywhere. The NLP layer reads only this.

What this is, and what it isn't

Warehouse-agnostic

We don't replace Snowflake, Databricks, or Fabric — we sit on top. If you already invested in a warehouse, we ride that investment.

Governed, not freelance

The LLM doesn't invent metrics. It can only reference governed definitions in the semantic layer. Every number is traceable to a definition, a SQL query, and the DQ checks that ran.

More than a chatbot

Same engine powers ad-hoc Q&A, scheduled digests, anomaly detection, and KPI exploration for executives still defining what to measure.

Trust is the product

Every answer ships with its lineage. If trust breaks, the program dies — so trust is a first-class feature, not an afterthought.