Selected Work

Products, not slideware.

A decade of taking AI from idea to production — agentic systems, conversational AI, and applied ML. Each project below is framed by the problem, what I built, and the measurable result.

10+
years in product
5
industries shipped in
15+
AI/ML products to production
$B+
in transactions touched by my products
Download résuméStart a conversation
Try a live demo — agentic task runner

Run a real agent loop with tool use and a human-approval gate. No sign-up.

Open demo
WEXLead PM, GenAI2024–25

Finn — Enterprise Multi-Agent Platform

Problem. Knowledge workers lost hours each week stitching together data across a dozen internal systems and chasing routine approvals. Existing chatbots could answer questions but couldn't take action, so adoption stalled after the novelty wore off.

What I did.
  • Framed the product around 'jobs that finish themselves' — agents that complete a task end-to-end, not just retrieve an answer.
  • Designed an orchestrator + specialist-agent architecture with tool-use, human-in-the-loop approval gates, and per-action audit logging.
  • Shipped a thin, observable v1 (3 high-frequency workflows) and instrumented every step to learn where agents failed or needed a human.
  • Built an evaluation harness so each new agent was graded on task success, not vibes, before release.

I owned product strategy, the agent UX, and the evaluation framework, partnering with an ML platform team of ~6 engineers.

LangChainVertex AIPythonTool-use / MCPVector search

Impact

~50%
less manual effort on automated workflows (vs. pre-launch baseline)
3 → 15
production workflows in two quarters
24×7
availability with human approval on sensitive actions
WEXProduct Lead2023–24

Virtual Agent for Call-Center Deflection

Problem. High call volume on routine requests (card replacement, plan changes) drove cost and long hold times, while live agents were stuck on repetitive work instead of complex cases.

What I did.
  • Mapped the top call drivers and prioritized the 20% of intents that drove ~70% of volume.
  • Designed voice + chat flows with safe fallback to a human the moment confidence dropped.
  • Integrated identity, card, and plan systems so the agent could actually complete the transaction, not just describe it.
  • Ran a contained pilot, watched transcripts weekly, and tuned intents before scaling.

I led discovery, flow design, and the rollout plan; partnered with conversation-design and platform engineering.

Dialogflow CXTwilioVertex AIUiPathMCP

Impact

~30%
reduction in routine call volume
Higher
live-agent capacity for complex cases
<1 day
card-replacement self-service turnaround
T-MobileSenior PM2021–22

Self-Install Experience with AR

Problem. Home-internet self-installs generated avoidable support calls and truck rolls when customers placed the gateway poorly, hurting both NPS and margin.

What I did.
  • Used AR to visualize optimal router placement and signal strength in the customer's actual room.
  • Designed a step-by-step flow with real-time validation so users knew immediately if a step worked.
  • Closed the loop with diagnostics that pre-empted the most common install failures.

I owned the install product surface and worked with AR engineering and care analytics.

ARKitReact NativeTelemetry / diagnostics

Impact

+25%
install-flow NPS
−40%
install-related issues
Fewer
support contacts and truck rolls

More work

WEXConversational AI

Benefits Assist

AI assistant that guides members through benefits questions and administration.

Faster
benefits resolution & lower support load
PythonRAGSQL
WEXApplied ML

Late-Fee Forecasting Model

Predicts late-payment probability to focus collections and reduce write-offs.

~98%
validation accuracy on held-out data
XGBoostDataiku
WEXApplied ML

Attrition Prediction

Flags at-risk accounts early so retention can intervene before churn.

Earlier
churn signals for proactive save offers
Pythonscikit-learnPandas
T-MobileConversational AI

Conversational AI Product Suite

Unified virtual-care experience spanning retail and digital channels.

Unified
care across retail + digital
DialogflowCustom NLU
EYApplied ML

AI-Powered Tax Technology

NLP + RPA tooling to streamline tax processing for global enterprises.

−35%
processing time, higher compliance accuracy
PythonNLPRPA
HiltiPlatform / IoT

Connected Tools Platform

Tracking, analytics, and management for industrial tools at scale.

−20%
tool loss rate, better asset visibility
AngularIoT SDKsAzure

A note on numbers: figures are rounded and, where work is under NDA, anonymized. Each reflects outcomes I directly drove with my team — happy to walk through the details and my specific role in a conversation.