Generative-AI product · Real-time NLP · Capital One
I lead and hands-on built a real-time, agentic summarization system that drafts a contact-center agent's post-call notes as the conversation happens — turning messy transcripts into clean, structured records.
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 problem
Contact-center agents carry a per-call administrative load: post-call notes written in heavy shorthand that is hard to use for quality assurance and audit — and agents sometimes pause a call just to finish notes and protect their after-call-work time. The opportunity: model-drafted, plain-text summaries that are simultaneously useful to agents, to QA, and to the downstream systems that consume call records, with real upside for compliance in call-result logging.
What I did, end to end
I owned this from roadmap to production: scoped timelines and what “done” meant, drove the model-performance analytics with a research partner team, and worked the months-long, bank-grade model-risk governance process that any customer-adjacent generative system has to clear. Then I built it hands-on — an LLM-backed summarization service plus the streaming glue that delivers a finished summary onto the agent's screen.
Part of the story I'm proud of: the build ran through an orchestration stack I'd never used. I came up to speed in days using agentic coding (Claude Code) plus disciplined self-review, wrote the full pipeline, and shipped it with a partner engineer reviewing my changes — final shadow runs landing the evening before the pilot.
How I measured it
I validated summary quality before pilot with a two-tier eval harness — human annotation plus an LLM auto-judge scoring factuality and completeness — and agents accept a high share of drafts with no edits. The operational upside is framed as projected after-call-work savings at full rollout, using a transparent wage-based cost-per-minute methodology; I quote it as a projection, not realized savings. It's live in a phased pilot.
What it demonstrates
Carrying an ambiguous, cross-functional GenAI product from roadmap through governance to production; evaluating a non-deterministic system honestly before trusting it; and using agentic coding to move fast on unfamiliar infrastructure without cutting the review corners that production demands.