Work
35+ production AI deployments across 10+ industries. Most I built or ran; the ones I advised on are marked Advise.
One Virtual Agent Platform, Every Channel, One Telco
12M+ subscribers · $350M+ OPEX relief
My role: Chief Architect & Product Owner · 50 people across North America, Latin America and Europe
The Challenge
Three regions, each with its own telephony and CRM, and about 100 backend endpoints (REST, SOAP, gRPC and direct database calls) that the agents had to act through at low latency.
What I did
IBM North America's first contact center AI program, and I ran it for eight years. My team built seven virtual agents for subscribers, field technicians and employees across chat, voice, SMS and the app, on an orchestration layer we wrote from scratch. IBM was paid per deflected interaction. We re-platformed it twice without dropping traffic. It still runs.
Scale
12M+ subscribers
Copilot Studio Agent Factory
800+ ungoverned agents found and remediated
My role: Platform Owner & Technical Program Manager · 33K employees
The Challenge
Letting non-technical employees build thousands of agents without losing track of what each one can read, write or spend.
What I did
I stood up the Agent Factory behind the client's Copilot Studio rollout with IT, Security, HR and engineering leads, none of whom report to me. It runs on an agent registry, graduated maker permissions, connector-tiered red teaming and an eval harness that retires pilots. With the Copilot Studio Kit we found 800+ agents built without oversight, traced each to its maker and cleaned them up app by app. I also wrote the ETL that prices each agent's credit use, which leadership now uses as its AI ROI baseline.
Scale
Global enterprise
Unified Employee Concierge
Copilot Studio vs EmployeeWorks vs Semantic Kernel
My role: Forward Deployed Engineering Lead
The Challenge
Making sure a user's permissions survive every agent hop and context carries from one agent to the next, without creating a Frankenstein UX.
What I did
I'm leading the build of one employee front door three ways, so the client can choose on data: low-code on Copilot Studio, off-the-shelf on ServiceNow EmployeeWorks, and custom on Semantic Kernel with Azure AI Search. It reaches Now Assist and Workday over A2A and other systems of record over MCP.
Scale
Global enterprise
AI Platform Bake-off & AI Factory
Glean vs Sana vs custom Vertex AI · ~5TB
My role: Forward Deployed Architect & Delivery Lead
The Challenge
A firm whose product is advice had to pick an AI platform its own consultants would actually use.
What I did
I ran the platform bake-off comparing Glean, Sana and a custom Vertex AI search across ~5TB of multimodal content, scored on cost and quality. The firm had nearly given up on enterprise search. The CIO took Glean firm-wide, and about 5,000 consultants use it. I'm now setting the architecture for the firm's AI Factory across $500M+ in identified value and building a persona-based eval and tokenomics workbench across its AI platforms.
Scale
Global consulting firm
Agentic HR & IT Self-Service
40% containment · 30-point CSAT gain
My role: Chief Architect & Delivery Executive · 15 people
The Challenge
Organizational RACI alignment between historically split HR and IT business units, plus strict cloud infrastructure requirements for a regulated utility.
What I did
HR and IT self-service for 23K employees on ServiceNow Now Assist, with agent assist for the help desk and auto-written resolution notes. It was the client's first agentic AI in production. I brought Claude into ServiceNow through Amazon Bedrock with AWS, which co-invested in the multi-year production run, and held every release to a routing-F1, RAGAS and LLM-as-judge bar.
Scale
23K employees; multi-$M deal
GenAI Open Enrollment Assistant
70% containment · ~$1M/yr saved
My role: Chief Architect
The Challenge
Strict private endpoints, complex private network topology, encryption at rest and in transit — in a regulated environment with a compressed launch window.
What I did
IBM Consulting's first generative AI solution in production: an HR assistant for open enrollment that my team owned end to end, from architecture and build through launch and change management. We started in early 2024 with no playbook. The design used a mixture of experts, prompt generation and LLM-as-judge evals, with Azure OpenAI and AI Search behind private endpoints to cut latency and protect HR data.
Scale
Enterprise benefits population
AskHR, IBM's HR Assistant
10M+ conversations a year · .98 routing F1
My role: AI Product & Architecture Lead
The Challenge
Routing every HR question from hundreds of thousands of employees to the right answer, in a domain where a wrong policy answer has consequences.
What I did
IBM's own HR assistant, used across the company for more than 10 million employee conversations a year. I raised first-try intent recognition 20 points. Then I moved it to a transformer-based master and child agent design (T5, RoBERTa) at .98 routing F1. I later led its GenAI replatform onto Watson Orchestrate, a new IBM product line shaped partly by what we learned on AskHR.
Scale
IBM global workforce
7M-Subscriber AI Contact Center
7M subscribers · 40% containment at launch
My role: Chief Architect
The Challenge
Building a greenfield AI contact center under aggressive timelines during a corporate spin-off — new brand, new org, inherited tech debt.
What I did
Stood up the contact center from nothing during a corporate spin-off, with voice and digital channels live at launch. Then moved it from Salesforce and Watson to Google CCAI and Dialogflow while the parent company was still separating systems.
Scale
7M subscribers
Conversational Insights
Built for one client, sold to 20+, folded into Watson Assistant
The Challenge
Processing tens of millions of conversation records with mixed quality and delivering actionable insights fast enough to influence weekly product decisions.
What I did
An ETL and transformer-model pipeline that found the friction points in every virtual agent conversation, years before assistant platforms shipped their own analytics. It started at the telco, sold to 20+ Watson clients, and IBM later built it into Watson Assistant.
Scale
Tens of millions of records
AIOps Noise Reduction
43% less alert noise · ~$2M/yr saved
The Challenge
Hundreds of millions of alerts across 200K devices with inconsistent signal quality — alert fatigue was slowing incident response to a crawl.
What I did
Built a real-time telemetry pipeline that streams alerts from ~200K network devices through an XGBoost classifier, so NOC and field teams see actionable incidents instead of false positives.
Scale
200K+ devices
Copilot Studio HR Assistant
~28K employees · Copilot Studio + SAP Joule
My role: Technical Program Manager
The Challenge
An inherited build for a largely union-represented workforce, where what the assistant can and cannot answer had to be designed explicitly.
What I did
Leading the build of an HR assistant for ~28K employees, about 18K of them union-represented. First I'm fixing the inherited build (topic structure, agent instructions, evals, observability, search indexes), then adding actions and A2A with SAP Joule.
Scale
~28K employees
Enterprise AI Academy
Enterprise-wide GenAI literacy & adoption
The Challenge
Running a structured AI literacy program while the GenAI landscape evolved weekly — balancing foundational knowledge with cutting-edge tool adoption.
What I did
Designed and ran a 6-month AI literacy program for a large media and telecom company, with office hours, a capstone project and a use-case pipeline tied to governance.
Scale
Enterprise-wide
Agentic HR Service Delivery
100K+ employees · Fortune 20 scale
The Challenge
Deploying agentic AI in a highly regulated healthcare environment with strict PHI/PII controls and a workforce spanning clinical and corporate roles.
What I did
Product owner for an HR and IT employee assistant for 100K+ employees, built on the client's own harness over Azure OpenAI because Now Assist didn't fit their architecture. I designed how ServiceNow knowledge articles, catalog items and topics were indexed into Azure.
Scale
100K+ employees
B2B Sales Agent Assist
Real-time sales assist for enterprise DIA
The Challenge
Sales reps needed instant access to complex B2B product configurations and pricing across a fragmented knowledge base during live customer calls.
What I did
Built an agent assist tool for enterprise DIA sales reps that surfaces technical specs, pricing guidance and competitive positioning in real time during customer calls.
Scale
Enterprise B2B sales
AI Center of Excellence
Multi-year AI operating model · Multi-$M
The Challenge
Sustaining strategic alignment and delivery momentum across a multi-year engagement spanning enterprise product teams.
What I did
Led a multi-year AI Center of Excellence for the telco's enterprise product teams. We ran the use-case pipeline and vendor bake-offs, and my team built what came out of them.
Scale
Enterprise-wide; multi-$M contract
Field Service Virtual Agents
Reduced field troubleshooting time
The Challenge
Field techs needed hands-free, step-by-step troubleshooting guidance while working on live network equipment — latency and accuracy were non-negotiable.
What I did
Deployed virtual agents that walk field technicians through troubleshooting, parts lookup and dispatch, integrated with Nokia network management.
Scale
National field workforce
Global HR Assistant
300K+ employees · Global deployment
The Challenge
Scaling a single conversational AI across 300K employees in a partnership-driven culture with complex, region-specific HR policies.
What I did
Led delivery of a global HR virtual assistant for 300,000+ employees across regions. It answers benefits, policy, onboarding and HR service questions in chat and collaboration channels.
Scale
300K+ employees
Predictive Sales Assistant
Real-time intent scoring for sales intervention
The Challenge
Translating noisy clickstream data into reliable purchase-intent signals fast enough to intervene during the browsing session.
What I did
Built a proactive sales assistant that analyzed real-time web browsing behavior to predict purchase intent and trigger contextual chat interventions.
Scale
Consumer web traffic
Audit & Underwriting AI
Accelerated Big Four audit review cycles
The Challenge
Extracting structured insights from massive, unstructured audit evidence packages where accuracy directly impacts regulatory compliance.
What I did
Worked on IBM's joint initiative with a Big Four audit firm to apply Watson to audit and underwriting, with document understanding over large, complex evidence sets.
Scale
Big Four audit practice
Autonomous Ride Assistant
<2% failure rate · TechCrunch, CNET coverage
The Challenge
Building reliable voice interaction in a moving vehicle with ambient noise, short response windows, and live public demonstrations with zero margin for error.
What I did
Built the in-vehicle voice assistant for Olli, the first 3D-printed autonomous shuttle: dialog management, speech in and out, and live vehicle telemetry. It was shown at CES and IFA Berlin.
Scale
Global press coverage
Parts Intelligence Platform
Automated competitive pricing intelligence
The Challenge
Matching unstructured parts descriptions across manufacturers with different naming conventions, then building reliable price comparisons across fragmented data.
What I did
Built an intelligent parts analysis system that identifies competitive aftermarket alternatives and performs automated price benchmarking across the parts catalog.
Scale
Global parts catalog
Platform Migration
Zero-downtime platform migration
The Challenge
Migrating a live, high-volume engagement platform with zero downtime tolerance.
What I did
Led the migration of the customer engagement platform from LivePerson to Salesforce. Conversation flows, routing and AI integrations were re-architected while the service stayed live.
Scale
12M+ subscriber platform
SMS Channel Deployment
AI contact center extended to SMS
The Challenge
SMS supports only plain text — no rich media, no carousels, no buttons. Delivering a sophisticated experience required creative conversation design.
What I did
Deployed the virtual agents over SMS through Salesforce. SMS is plain text only, so every flow was redesigned without buttons, carousels or rich media.
Scale
National SMS channel
Fab Technician Assistant
Reduced fab troubleshooting time
The Challenge
Translating highly technical semiconductor process documentation into reliable, step-by-step conversational guidance.
What I did
Built a conversational assistant that guides semiconductor fab technicians step by step through troubleshooting complex chip fabrication equipment.
Scale
Fab technician workforce
GenAI HR Assistant
GenAI HR on Bedrock + ServiceNow
The Challenge
Integrating a GenAI assistant with ServiceNow in a regulated environment with strict private endpoint requirements.
What I did
Built an HR self-service assistant on Amazon Bedrock and connected it to ServiceNow HRSD, inside a regulated utility's private-endpoint requirements.
Scale
Utility workforce
Multilingual Agentic HR
30K employees · Multilingual agentic HR
The Challenge
Serving employees across multiple countries with different languages, labor laws, and cultural expectations from a single agentic system.
What I did
Leading the design of multilingual agentic HR service for a 30K-employee multinational, with region-specific compliance in each language.
Scale
30K multinational employees
Transactional HR Agent
Q&A to transactional HR agent
The Challenge
Moving from "answering questions" to "completing transactions" required deep integration and careful guardrails around irreversible actions.
What I did
Took the HR virtual agent from answering questions to completing transactions, so employees can finish HR actions inside the conversation.
Scale
National telecom workforce
AI-Powered Site Search
Semantic search for millions of visitors
The Challenge
The site served millions of visitors across consumer, small business, and enterprise — search had to understand intent across wildly different journeys.
What I did
Replaced the telco website's keyword search with semantic search and contextual ranking on Watson Discovery.
Scale
Millions of web visitors
SOX Compliance Automation
70% SOX review time reduction
The Challenge
SOX reviews are labor-intensive manual comparisons across hundreds of apps — but automation must be explainable and auditable.
What I did
Applied agentic AI to SOX compliance review. The agent reads admin screenshots across hundreds of applications and flags anomalies, which cut manual review time 70%.
Scale
Hundreds of applications
Credit Scoring Model
Improved B2B credit decisioning
The Challenge
Traditional credit scores don't capture telecom-specific risk signals — needed to blend traditional and alternative data without introducing bias.
What I did
Built an auxiliary credit scoring model for mid-market enterprise customers that adds behavioral and account signals to traditional scores.
Scale
Enterprise B2B credit
Harvest Yield MLOps
Production MLOps for seasonal prediction
The Challenge
Building reliable ML infrastructure for agricultural prediction where ground truth is seasonal and the business gets one harvest cycle per year to validate.
What I did
Defining the MLOps architecture for a harvest-yield prediction system — satellite imagery pipelines, model training infrastructure, and seasonal forecasting.
Scale
Agricultural operations
Drive-Thru Voice Agent
Stable voice ordering in high-noise environments
The Challenge
The rigor of production-grade NLU in the wild — accurately transcribing a menu order while a baby cries in the back of the car.
What I did
Advised on the design and build of an automated order-taking voice assistant for drive-thru lanes. It had to handle menu complexity, ambient noise and real-time POS integration outdoors.
Scale
National QSR chain
Member Contact Center AI
45M+ members · Fortune 20 architecture
The Challenge
Navigating healthcare regulatory requirements while designing conversational AI that handles sensitive health insurance inquiries with empathy and accuracy.
What I did
Advised on the architecture for member-facing contact center automation at a Fortune 20 health insurer with 45M+ members across voice and digital channels.
Scale
45M+ members
Citizen Support Agent
1.6M citizens · State government
The Challenge
Designing citizen-facing AI for government with strict accessibility requirements and zero tolerance for misinformation.
What I did
Defined the architecture for a citizen-facing support agent for 1.6M residents that answers eligibility questions, routes services and opens cases.
Scale
1.6M citizens
HR & ITSM Virtual Agent
Greenfield AI for post-spin-off company
The Challenge
Standing up AI service delivery from scratch for a newly spun-off company with inherited processes but no AI practice.
What I did
Advised on the architecture and delivery of HR and ITSM virtual agents on ServiceNow for a newly independent consumer health company with no AI service delivery in place.
Scale
Global workforce
Intelligent Contract Review
Automated contract redlining for legal
The Challenge
Legal teams have zero tolerance for hallucination — building AI that lawyers trust required explainability and careful confidence thresholds.
What I did
Contract analysis on Azure OpenAI for the legal team: auto-redlining, deviation detection and clause generation.
Scale
Legal department