Real-time NLP · MLOps · Capital One
This is my formative project: the first real-time model in the company's call centers — live sentiment that flags a frustrated caller within about half a second so a supervisor can step in. It started as an annotation exercise and now runs across the business.
This is a deliberately high-level account of proprietary work at Capital One — the problem, my role, and the approach, described qualitatively. It contains no confidential figures, internal systems, data, models, code, vendor or team names, or architecture.
The first-of-its-kind bet
When real-time transcription first became possible internally, I joined a newly stood-up effort to alert on frustrated callers for escalation — at a time when the only models in the call center were batch or whole-call. Doing it live, per utterance, was new ground.
The data foundation
To make modeling possible I organized a multi-week annotation program: individual utterances labeled for sentiment across thousands of calls, with regular inter-annotator-agreement checks to keep the labels honest. Then weeks of modeling narrowed the approach to the best call-level precision/recall — and a transformer model fast enough to serve in real time.
Serving, scale, and carrying it solo
Working hand-in-hand with a partner platform team, I tuned the deployment to real-time inference latency, with a real-time end-to-end path from speech to on-screen alert. When the effort lost its resourcing, I carried the pilot on my own: I built a deliberately simple per-team alerting path in the tools agents already used — standing up a new team quickly with no engineering effort — tracked the pilots, and built the performance analytics for senior leadership.
Where it is today
The model now runs across the business at production scale, and the scrappy pilot path has been replaced by dedicated integrations in agent-facing UIs. Real-time modeling in the call center started here — and the next wave I'm building tags every call with dozens of QA, alert, and extraction signals.