Agentic GenAI Assistant by Miguel Marinho

Summary

Miguel Marinho leads a production-grade AI agent on Azure Databricks that turns supply-chain business requests into governed AI workflows. The system gives teams one natural-language interface for grounded document answers, live SQL analysis, and shareable business reports.

Problem

Supply-chain teams needed one governed assistant for both knowledge retrieval and analytics: trustworthy answers from SharePoint documents, safe access to live Databricks data, and a faster path from business questions to usable reports.

My role

Miguel leads the architecture, delivery, and deployment of the agent, spanning retrieval, Text-to-SQL, natural-language reporting, human-in-the-loop evaluation, LLMOps, least-privilege access, and Azure DevOps CI/CD across SIT, QA, and Prod.

Architecture

• Knowledge Agent: retrieval-augmented Q&A over curated internal documents, ingested, chunked, and indexed with Mosaic AI Vector Search, returning grounded answers with in-context document images.
• Text-to-SQL Agent: users ask business questions in plain English; the agent retrieves table and column metadata through vector search, generates Databricks SQL, and executes it against a SQL Warehouse through the Statement Execution API.
• NL-to-Report Generator: users describe a report in plain language; the agent produces an API-consumable config that a deterministic service compiles into SQL, executes, and serves as a shareable report.
• Human-in-the-loop evaluation: thumbs-up/down feedback is logged and surfaced in Databricks dashboards for systematic review and targeted quality improvement.
• Orchestration and LLMOps: agentic routing, evaluation, tracing, lineage, custom deployment, least-privilege access, and CI/CD across SIT, QA, and Prod.
• AI-augmented development: Cursor agentic coding workflows with repo-aware context, custom rules, SKILL.md procedures, and MCP-connected internal tools.

Outcomes

• Supply-chain users can move from question to answer, query, or report without jumping between documents, SQL tools, and self-service reporting workflows.
• The assistant preserves traceability and governance through grounded retrieval, deterministic reporting services, Databricks controls, and least-privilege service principals.
• The project established a production-ready agent architecture spanning retrieval, analytics, reporting, evaluation, CI/CD, and continuous improvement from user feedback.