Skip to content
← cd ~/projects

Plain-English questions, answered with real SQL

Agentic analytics bot

Built an agentic analytics bot: 14 typed DuckDB-backed tools querying 2,400+ outlets via AWS Bedrock, helping drive personalised outlet improvements.

14
typed agent tools
2,400+
outlets covered
5
tool rounds max

(Python · AWS Bedrock · Claude · DuckDB · SQL)

  1. # Problem

    Operators needed answers from outlet-level data without writing SQL. A free-form LLM over the warehouse would make up filters, tables, and numbers.

  2. # Approach

    Shipped an agent on AWS Bedrock (Claude) with 14 typed tools, each backed by a parameterised DuckDB query. A tool-use loop of up to 5 rounds lets the model gather what it needs, and routing guards keep questions on supported paths, so every answer comes from a real query rather than model memory.

  3. # Result

    Plain-English analytics across 2,400+ outlets, used to steer outlet-specific improvements, with every number traceable to a query.