AI Strategy & Architecture · Agentic Systems · Production AI
Paul Henkelman
Paul Henkelman is an enterprise AI executive who designs and delivers AI systems that operate under real production conditions: agentic services in customers' hands, AIOps platforms watching a 30M+ subscriber network, and the training and inference infrastructure underneath them.
Sr. Director, AI Strategy & Architecture at Charter Communications. Invited speaker at CableLabs and SCTE. Author of Noetix, an open-source agentic memory system.
Core Focus
Most of the work lands in one of four areas.
Architected Charter's first customer-facing agentic AI service, presented to leadership up to the CEO and adopted as the company-wide reference architecture; deep protocol-level MCP work, guardrail and human-in-the-loop patterns, and agent evaluation.
Idea-to-inference platform design (GPUaaS), distributed training and inference at scale, and 100+ LLM fine-tuning runs personally directed on EKS and SageMaker, through full pretraining and world-model training.
Detection, causal inference, prediction, and automated remediation for a 30M+ subscriber footprint; petabyte-scale telemetry ML built on a decade pioneering enterprise AIOps at Comcast.
Enterprise AI strategy, Data & AI governance on Charter's AI Center of Practice, and AI education at scale as author and instructor of Charter's internal AI curriculum for engineers and executives.
Systems
Agentic services in customers' hands, and the platforms watching a national network.
Charter's first customer-facing agentic AI service
Conceived the solution, built the fully functional prototype, and designed the production architecture (LangGraph, LangChain, custom and standardized MCP, Playwright-MCP) for an agentic system that executes service-activation steps on the customer's behalf, reducing roughly 2 hours of customer effort to about 2 minutes.
The demo was presented at every level of leadership, including to the CEO; guided the cross-team build to production launch serving thousands of customers per day. The design was adopted as Charter's reference architecture governing ongoing agentic development.
Detection, causal inference, prediction, and automated remediation · 30M+ subscribers
Designed the detection, causal-inference, prediction, and automated-remediation platform (Random Cut Forest, MAD-GAN, Bayesian networks) for customer-facing network issues across a 30M+ subscriber footprint; delivered the full PoC and implementation design and led multi-department development at enterprise scale.
Builds on pioneering Comcast's enterprise AIOps platform: ML alerting and event-correlation pipelines that cut false-positive alerts by 92%, petabyte-scale telemetry architecture, and automated anomaly response that materially reduced MTTR.
Idea-to-inference platform on edge GPU infrastructure
Lead inference and MLOps/LLMOps architecture for a GPU-leasing initiative: an idea-to-inference platform with continuous training and enhancement loops, built to serve external AI workloads at commercial scale.
Personally direct large-scale training on EKS and SageMaker: 100+ LLM fine-tuning runs across Qwen and other open-weight families, full LLM pretraining, and world-model training.
Speaking
Three invited talks in 2026: enterprise and agentic AI, AI & data governance, and world modeling.
April 2026
AI & Data Governance
June 2026
Agentic AI and Enterprise AI
July 15, 2026
AI in Network Management & Control, and the World-Model Future
Writing
How these systems actually work, and which design choices stand up in production.
Essay · Aug 7, 2026
A complete transformer, traced by hand: tokens, embeddings, attention, detectors, and the final prediction, every number on the table. If you can multiply and add, you can follow all of it.
Read essay
Essay · Mar 9, 2026
Why useful AI agents need more than just tools: the integration argument everyone is having is downstream of an architecture question almost nobody is asking.
Read on Medium
About
Thirty years across systems, networking, security, and software, the last decade leading AI at Fortune 100 scale. What stuck is a habit of judging architecture by how it behaves under operational load.
For conversations on AI architecture, agentic systems, speaking, or the operational realities of large-scale AI platforms.