4

Conversational Analytics POC — Global Beauty Conglomerate

Self-service conversational analytics (NL → SQL → answer + charts) for a leading global enterprise in a regulated sector; LangGraph multi-agent pipeline; FastAPI + Next.js; SQLite/PostgreSQL.

Self-service conversational analytics delivered for a leading global enterprise (beauty conglomerate) in a regulated sector. Non-technical users ask questions in natural language and receive instant answers with charts, reducing dependency on IT and shortening time-to-insight.

Purpose and Goals

  • Business: Direct access to enterprise data without writing reports or waiting on IT; natural-language questions and instant answers with charts.
  • Technical: End-to-end flow: natural-language question → intent routing → SQL generation and execution → formatted answer with optional auto-selected visualisations; conversation history and optional external context.

Challenges

  • Semantic–database alignment: Mapping casual user phrasing to canonical values required context-aware entity resolution; step-based categorical matching so intent maps to the right columns and values.
  • SQL reliability: Safe, correct SQL across complex schemas with strict escaping, validation, and clear fallbacks when generation failed.
  • Concurrency and UX: Async request manager with configurable timeouts, queue limits, and request cancellation by ID.
  • Multi-backend support: Same pipeline for SQLite (development/demos) and PostgreSQL (including managed cloud), with health checks and consistent connection handling.
  • Pipeline complexity: Coordinating multiple agents (router, clarification, SQL generation/execution, analysis) with a clear state model and consistent error handling.
  • Deployment: Docker to cloud with secrets management, environment-based configuration, and correct frontend–backend networking.

Deliverables

  • Conversational analytics: Full flow from natural-language question to text answer and optional visualisation; line, bar, horizontal bar, scatter, and pie charts with automatic chart-type selection and short reasoning explanation.
  • Multi-agent architecture: LangGraph-based pipeline with routing, clarification, SQL generation and execution, and analysis agents; streaming where applicable.
  • Backend: FastAPI with async request manager (semaphore, queue, configurable timeout, cancel-by-UUID), Bearer-token auth (Auth0-ready), CORS, health checks for API and database.
  • Frontend: Next.js chat application with conversation history, saved and preset questions, internationalisation (EN, PL, FR), theme support, responsive layout.
  • Data and deployment: SQLite and PostgreSQL for analytics; SQLite for chat persistence; Docker Compose and cloud deployment with parameterised configuration.
  • Quality: Evaluation datasets and test suites for agents and API (integration, cancellation, negative paths); structured architecture documentation.

Outcome

Working proof of concept demonstrating feasible, secure, natural-language access to analytics; foundation for scaling to more users and data sources and for faster time-to-insight. Reusable patterns: request queue, multi-agent graph, categorical matching, visualisation pipeline; clear path from POC to production.

The same platform was later delivered for a post-lease automotive trading client (see Post-Lease Automotive Platform).