Enterprise AI Platform Architect & Engineering Leader
Building reliable, secure, and scalable AI platforms — from enterprise infrastructure to agentic AI.
18+ years architecting and leading enterprise AI and data platforms at Mastercard, most recently as Director, AI Center of Excellence — supporting 2,000+ users and mission-critical analytics at global scale. I pair that with hands-on engineering in generative AI: Model Context Protocol (MCP) servers, LLM tool calling, OpenTelemetry, and AI evaluation.
Current focus: agentic AI architecture, AI observability, AI security, cloud AI platforms, and production-ready AI engineering.
Background
18+ years scaling AI and data platforms in Fortune 500 environments — most recently leading a global engineering team at Mastercard on enterprise AI infrastructure, including Kubernetes-based ML platforms, Cloudera AI, and internal LLM tooling. Deep roots in Kubernetes, distributed systems, platform engineering, and enterprise security.
Director, AI Center of Excellence / Platform Engineering2021–2026
Mastercard
Led a global team of 11 engineers supporting 2,000+ users; migrated 6,000 notebooks across 1,000+ servers to Cloudera AI at 99.9% uptime, moved key workloads to Kubernetes-based ML infrastructure (MTTR down 60%), and built an AI-powered support chatbot that cut ticket volume 30%.
Principal Architect, Data Science & AI Platforms2013–2021
Mastercard
Built the org's first enterprise open-source Python distribution (100+ libraries, PCI-compliant) and enabled fraud-detection analytics processing 143B+ transactions annually via graph database technologies.
Lead Developer → Senior Business Analyst → Senior Developer2007–2013
Mastercard
Drove Hadoop modernization scaling beyond 2 petabytes and led enterprise reporting migration off Business Objects, cutting platform costs 33%.
Full history, including earlier roles at PepsiCo and education (MBA, BS Computer Science), is in the resume.
Featured project: SatornTCG AI Assistant
An MCP server that lets an AI assistant work against the live data of a real trading-card resale business, built with the observability, testing, and access controls I'd expect of an enterprise platform.
In daily use
satorntcg-mcpserver
The problem
Running a resale business means constant questions: what's underpriced, which box paid off, which orders still need listings. Answering them meant clicking through dashboards. This server exposes the business's inventory, pricing, P&L, and order data as standardized tools, so an assistant like Claude can answer from live data instead of a demo set.
Architecture
AI AssistantClaude Desktop / Code, or a custom CLI agent
→
MCP ServerNode.js, tool discovery & execution over stdio
OpenTelemetry across every hop → distributed traces in Jaeger · agent metrics in Prometheus / Grafana
Key engineering
MCP tool layer — 11 tools covering inventory, pricing, box and business P&L, eBay listings, TCGplayer order sync, and eBay deal scanning, each published with a JSON input schema for discovery and execution.
Validated, bounded queries — server-side argument validation with filtering, sorting, and hard row limits, since the low-level MCP server doesn't enforce schemas itself.
Agent runner with guardrails — a CLI agent on the Claude API that offers only read-only tools, checks every argument against the tool's schema before calling it, and stops after 5 tool rounds.
Resilient integrations — bounded retries with jittered backoff for idempotent HTTP reads only. Write calls never retry.
Testing & observability
78 automated tests passing (Node test runner): argument validation, an MCP integration test that starts the real server and checks tool discovery, agent loop logic, retry behavior, and telemetry export.
LLM evaluations — a live Claude API eval checks tool selection and arguments for natural-language questions, and checks that the model answers directly, asks for clarification, or declines when no tool fits. Repeat runs measure consistency.
Distributed tracing — every tool call is a span with each database query as a child span, so slow calls are flagged automatically and you can see whether time went to the database or the handler. Trace context is passed from agent to server through MCP request metadata, so one trace covers the whole request.
Metrics — request, tool-call, latency, and token-usage metrics through an OpenTelemetry Collector into Prometheus and Grafana. Labels are kept to an allowlist, so prompts and data never end up in metric labels.
The commerce platform behind SatornTCG: inventory, price alerts, eBay/TCGplayer listing management, sales and P&L reporting, and a RAG-based rules assistant, on Postgres with row-level security.
Web chat interface — a hosted version of the assistant on the Claude API, replacing the local desktop client.
Remote HTTP transport with authentication, alongside today's local stdio transport.
Hosted tracing backend and Grafana dashboards for the agent metrics.
Larger eval set covering more tools and adversarial prompts.
AI security & observability
An agent that can query and change business data is a platform like any other: it needs access control, monitoring, and regression testing. What's built and what's next:
Built in SatornTCG Implemented
Least-privilege tool access — the agent gets an allowlist of read-only tools, and any other tool request is rejected before it reaches the server.
Input validation at two layers: schema checks in the agent, plus validation and row limits on the server.
Data protection in telemetry — agent traces and metrics leave out prompts, message content, tool results, and credentials, and a test confirms retry telemetry never records headers.
Secrets handling — credentials stay in local environment files and out of source control. The web app uses Postgres row-level security across all its views.
Tracing and tool monitoring, with slow tool calls flagged automatically.
Evals as regression tests for tool selection and refusal behavior.
Studying & planned Not yet built
Human approval enforced on the server for tools that write data, rather than relying on the client's confirmation prompt.
Prompt-injection defenses for tool results that contain untrusted third-party text, such as marketplace listing titles.
Authentication and per-user authorization for remote MCP access.
Audit logging and governance — a durable record of who triggered which agent action, and why.
Skills
Grouped by where each skill comes from: years in production, current hands-on building, and recent cloud coursework.
Selected recent training. These are course completions unless marked as a credential.
Model Context Protocol (MCP): Hands-On with Agentic AI — put into practice in the SatornTCG MCP server above.
Building Generative AI Applications Using Amazon Bedrock · Amazon SageMaker AI Getting Started · AWS Cloud Practitioner Essentials (AWS Skill Builder courses)
Microsoft Certified: AI Business Professional (credential)
Academy Accreditation – Databricks Fundamentals · Generative AI for Business Leaders · Foundations of Responsible AI