All projects

PMG · Production AI platform

Ask Alli

A production AI platform at PMG for exploring analytics data and generating insights.

Overview

Ask Alli had grown into a large production AI application with workflows that were becoming harder to evolve and debug. The platform also needed better visibility into AI execution, a more modern frontend, and stronger safeguards against regressions as development accelerated.

My role

As one of two engineers leading development of Ask Alli, I work across the backend, AI orchestration, and frontend. My work includes designing LangGraph workflows, building reusable AI services, improving the React interface, and adding tooling that makes the platform easier to develop and operate.

Key contributions

  1. 01Led the migration from a monolithic backend to a modular LangGraph orchestration architecture with intent-based routing.
  2. 02Built reusable AI services on AWS that power automated insight generation across multiple PMG applications.
  3. 03Designed workflow nodes that surface trends, anomalies, top-performing segments, and cross-dimensional relationships from analytics data.
  4. 04Modernized the React frontend with real-time workflow status updates for long-running AI requests.
  5. 05Built PMG’s first automated API smoke-testing framework using Playwright and GitHub Actions to catch breaking changes before deployment.

Technologies

Node.jsTypeScriptReactLangGraphAWS ECSSQSS3DatadogPlaywrightGitHub Actions

Engineering challenges

Every change affects multiple layers of the product. Improvements to AI orchestration also need to work with APIs, frontend workflows, and production infrastructure. The challenge is making the platform more capable without sacrificing reliability or making it harder for other engineers to build on.

Lessons learned

Building AI products is as much about engineering systems as it is about language models.

Reliable orchestration, good observability, thoughtful evaluation, and fast feedback loops make it possible to improve AI behavior with confidence. Those pieces often have a bigger impact on the product than changing the model itself.

This case study is limited to public, non-confidential details about my work.