Intercom Fin Classifier

Support escalations contain early warning signals for production bugs, but individual tickets are scattered across support queues and buried under noise. The Intercom Fin Classifier transforms unorganized customer escalations into prioritized, actionable engineering defect clusters.

The Problem

When customers experience subtle edge-case bugs, individual support tickets trickle into Zendesk and Intercom over weeks or months. Because different customers describe symptoms with varying terminology, manual triage struggled to spot low-frequency patterns until dozens of users were impacted. One critical bug sat unaddressed for 107 days before reaching an engineering backlog.

What I Built

An automated defect discovery pipeline that extracts, vectorizes, and clusters support tickets without requiring pre-labeled training data:

1. Snowflake Ingestion Pipeline: Streams Zendesk and Intercom escalation logs directly into an embedding pipeline.
2. Vector Embeddings & Clustering: Computes OpenAI dense embeddings across ticket problem descriptions and runs density-based spatial clustering (DBSCAN) with cosine metric distance thresholds.
3. Automated Synthesis: For each discovered density cluster, an LLM synthesizes root cause hypotheses, customer impact summaries, and recommended engineering fix plans.
4. Toolchain Integration: Automatically maps clusters against existing Jira issues and Confluence architecture docs to avoid duplicate work.

Business & Engineering Impact

In backtesting and live evaluation across 26,000 historical escalations:

• Autonomously rediscovered 100% of open defects 15 to 75 days earlier than manual triage.
• Grouped noise into 20 actionable clusters, immediately highlighting 8 critical production bugs in 4 months.
• Demonstrated directly to the Chief Product Officer (CPO) and Product Platform leadership, who described the capability as holy-grail level functionality for voice-of-customer telemetry.