Hardened AI prototypes into production systems teams depended on daily. Turned early-stage agent prototypes into enterprise systems with reusable architecture patterns, robust error handling, CI/CD pipelines (GitHub Actions), nightly regression gates with 85%+ test coverage, and governance frameworks that teams could manage without engineering support. Won Indeed’s AI Innovation Hackathon (2025).
Built LLM conversational analytics and multi-agent pipelines, including a production system (Claude API + n8n) that compressed 40+ hours of manual analytics work to under 3 minutes across 4,794+ records with 91–92% coverage, driving a 30.8% operational improvement from baseline; designed RAG retrieval and guided analytics workflows blending traditional reporting with modern AI interaction.
Authored an open-source Tableau MCP Server (Model Context Protocol | OAuth 2.1 | Tableau REST, JavaScript, and Python SDKs) that lets LLM agents query, refresh, and reason over enterprise BI content, creating a reusable AI platform capability for data teams. Available on GitHub.
Engineered evaluation and observability for AI systems, including a five-axis LLM-as-judge evaluation harness (1.0 citation coverage) integrated with CI/CD via LangSmith, and LoRe, a self-healing LangGraph pipeline that detected and resolved 14 compliance drifts on its first production run. This made output trustworthiness a measurable, gated requirement rather than an assumption.
Owned the Tableau Next and Pulse product roadmap for Indeed’s enterprise analytics organization, supporting thousands of internal users; translated data scientist and analyst needs into platform priorities, piloted new AI capabilities with the Localization team before broader rollout, and drove adoption through documentation, reusable deployment kits, and hands-on enablement. Named a 2025 Tableau Ambassador.
Built analytics applications on a Snowflake production backend, including dbt semantic layers with governed metric definitions, multi-source integrations (Salesforce, Jira REST API, and Agiloft CLM), and a live production pipeline rebuilt with zero downtime.
