Not riding the AI wave —
I helped build it.

ZoomInfo had the data. It had the model. What it didn't have was adoption. Account Fit Score was at 30% when I joined — not because the model was wrong, but because the product experience was broken. Sales teams ignored it, routed around it, or used it wrong. I spent the first quarter talking to the people who weren't using it. The blockers were implementation complexity, opaque scoring logic, and a mismatch between model optimization and what reps needed to close deals.

I redesigned the explainability layer so reps could see why an account scored the way it did. I simplified onboarding. I built CRM feedback loops connecting win rates back into model refinement cycles. Twelve months later: 60% adoption. The product hadn't changed. The trust had.

From there I moved into agentic territory — LLM + RAG systems that autonomously segment accounts, generate outreach recommendations, and hand off to reps with full context. The key design challenge in agentic products isn't the automation. It's the handoff. I built explicit human-in-the-loop override workflows into every agent surface.

Moving into healthcare, I faced a domain where AI failure has the highest human cost. My mandate: build a document intelligence product that extracted structured data from unstructured clinical notes with HIPAA compliance built in from day one.

The hardest product decision wasn't the NLP architecture. It was defining what the system should refuse to do. I designed confidence thresholds below which Dexter escalated to human review rather than committing an extraction. I built audit trails into the core data model. We won 6 RFPs because our proposals showed we understood where the AI would fail — and had systematic plans for it.

This is where I learned what production AI really means. We built digital twin systems for PCB manufacturing — real-time computer vision pipelines that detected defects before they compounded and predicted equipment failures before they halted production. Leading a 25+ member team, I operated at every altitude simultaneously.

Clients saved over $1M annually — not because our model had the highest benchmark scores, but because we designed workflows around the model's uncertainty. Every critical decision had a human escalation path. That philosophy now applies to LLM outputs rather than CV model outputs — but the principle hasn't changed.

Before moving into product, I was the bridge between data science and engineering. I built annotation tooling, designed ML deployment pipelines, and watched models fail in ways no benchmark had predicted. This is where I learned what happens when a PRD is underspecified — and it permanently changed how I write them.

I was promoted directly from this role to Product Manager — an unusual path that gave me engineering credibility most PMs don't have when walking into ML design reviews.

BroadSoft powered UCaaS infrastructure for enterprise customers. Diagnosing failures meant parsing thousands of SIP signaling logs by hand. I built a diagnostic intelligence toolkit that indexed SIP log streams into Elasticsearch in real time, detected anomaly patterns, and surfaced root causes automatically. This was my first lesson in what data products do for operations teams — transforming reactive firefighting into proactive intelligence.

Burning Glass parsed millions of resumes and job postings to extract labour market insights. I wrote NLP extraction rules, managed Elasticsearch clusters, and improved a statistical model's accuracy from 95% to 98% through methodical rule engineering. This taught me that AI product work is 80% data quality and 20% model architecture — a belief I've carried ever since.