GenAI-Powered Voice of Customer Analytics
BERTopic to RAG — evolving support intelligence at scale at Autodesk
Story Behind the Project
It started with a casual conversation. During my time at Georgia Tech, I had worked on a system to route new StackOverflow posts to the right experts — getting questions answered quickly and correctly. When I mentioned this at Autodesk, it sparked a discussion: we should probably be looking at what people are posting on StackOverflow about Autodesk products.
That thread led us to the Developer Advocacy and Support (DAS) team, which handles customer support for Autodesk Platform Services. When we sat down with them, we found that they were dealing with support tickets coming in from multiple sources — emails, forms, and anonymous public posts pulled from StackOverflow tagged with #autodesk — with no automated way to make sense of the volume.
I ran an analytics deep-dive, built a POC for topic modeling, and pitched it to leadership. They saw the value immediately. What started as a single-team pilot eventually evolved into a modular platform, and then into a GenAI-powered system with RAG capabilities — a journey from BERTopic to retrieval-augmented generation over the life of the project.
About
A ML-powered customer support insights platform that proactively surfaces product issues and enables natural-language querying across 10,000+ quarterly support tickets. The system ingests tickets from multiple sources, applies topic modeling to identify emerging issue clusters, and was later extended with a RAG architecture for contextually relevant, open-ended querying beyond predefined topics. Originally piloted with one team, it was redesigned as a modular platform adopted by multiple product teams.
My Role
I led the project 0 to 1 — from the original analytics exploration and topic modeling POC, through leadership pitch and buy-in, to design, implementation, and platform expansion. I partnered with the Autodesk Platform Services team throughout, and drove both the technical evolution of the system (BERTopic → RAG) and the organizational expansion from a single-team pilot to a multi-team platform.
Key Details
- Piloted BERT-based topic modeling with the Autodesk Platform Services team; surfaced 7 critical product issues
- Extended system with RAG architecture layered on vector embeddings, FAISS, and cross-encoder re-ranking for open-ended natural-language querying
- Redesigned from a single-team POC into a modular platform with configurable data connectors
- Enabled 3 additional product teams to adopt automated ticket analysis with minimal engineering overhead
Impact
- 10.7% reduction in quarterly support ticket volume across 10+ product teams
- 7 critical product issues surfaced proactively via topic modeling; targeted fixes drove the volume reduction
- 3 additional product teams onboarded to automated ticket analysis via the modular platform redesign
- Natural-language querying across 10,000+ quarterly tickets via RAG
Technologies
Artifacts
The Georgia Tech project that sparked this work — a system to route new StackOverflow posts to the right experts for fast, accurate answers. Mentioning this in conversation at Autodesk is what led to the question: should we be looking at what people post about Autodesk products?
View on GitHub