I help teams turn ambiguous AI opportunities into practical systems with measurable business value.

Problem

Support knowledge creation was manual, slow, and reactive. Content lagged behind real customer needs, and the RAG experiences depending on that content could only be as good as what fed them.

What I built

Designed an end-to-end pipeline with specialized sub-agents for research, drafting, quality review, and fact checking, grounded against source material and iterated through review loops until the output cleared the bar. Extended it into a modular Knowledge AI capability layer covering gap detection, standards ingestion, generation, revision, and evaluation, combining classical ML and LLM methods where each fit best.

Impact

Cut article creation time by roughly 80 percent, from about a day to 40 minutes, with a pilot batch of AI-generated articles already deflecting tickets in production. The underlying capability stack and patent disclosures shaped roadmaps across adjacent product teams and shifted support from reactive triage toward proactive knowledge generation.

Problem

Support was spread across self-help, search, live chat, routing, and emerging AI experiences — a set of high-leverage surfaces with real cost and customer-experience stakes, but no shared foundation for measuring what was working or iterating with confidence.

What I built

Established instrumentation and an experimentation culture as the foundation, then iterated on the highest-leverage support surfaces — live chat, search and self-help, AI answers, and channel routing — grounding each change in close measurement of cost avoidance and customer-experience outcomes.

Impact

The layered work delivered $7M+ in documented cost avoidance and meaningful customer-experience gains — including a live chat routing improvement that shifted routed-to-chat share from 19% to 35% and a material reduction in time to resolution — turning the support experience into a system that could be tuned with evidence rather than guesswork.

Problem

A latent defect in PACCAR's MX-13 was projected to drive roughly $40M a year in warranty liability across a large field population, with failures surfacing late in a 5-year / 500K-mile window. The only fix was a $2,500 rebuild, and dealership capacity to perform it was tightly constrained.

What I built

Led a ten-person team across data engineering, ML, and warranty analysis, reporting weekly to the CEO and CFO. Built Cox proportional hazards survival models that scored each engine's 30-day failure probability from telematics duty-cycle features, plus a usage simulation that projected mileage and engine hours forward. The output drove monthly recall tranches, fed weekly root-cause reviews with the chief engineer, and powered a liability forecasting framework used to size warranty reserves on the P&L.

Impact

Saved $13M in warranty costs by pulling at-risk engines in for $2,500 rebuilds before they became $10K+ catastrophic failures, cutting unplanned customer downtime and getting more out of a constrained dealership network. The forecasting framework gave finance a defensible warranty reserve on the P&L, and the pattern of risk scoring under capacity constraints still shapes how I design AI systems today.