AI That Ships: The Operating Model and the Serving Stack

📢 Join DataPhilly for our October Tech Talks to learn what it takes to get AI working in the real world, from the operating model behind it to the way models are served.

Our host is Center for Business Analytics (CBA).
The CBA prepares students to become analytics leaders in their organizations. We bring together industry executives and world class faculty to create curricula and experiential programming that positions students to deploy analytics education and techniques to solve increasingly complex business problems.

Agenda:
6:00 - 6:30 PM: Arrival, networking, and refreshments
6:30 - 7:15 PM: Mariano Mattei: From Pilot to Production: The Operating Model Behind AI That Ships, followed by Q&A
7:15 - 8:00 PM: Vicki Boykis: Small but Mighty: ONNX + ONNXRT for Model Serving, followed by Q&A
8:00 - 8:30 PM: Networking

Speakers:
Mariano Mattei: From Pilot to Production: The Operating Model Behind AI That Ships
Abstract: Most large enterprises have already gotten AI into production: recent survey data puts the number above seventy percent. Far fewer can say the deployment is actually working. Half of those production-stage companies cannot prove ROI, and separate research shows the share of organizations converting AI activity into real bottom-line impact is closer to six percent. This talk argues that gap is not technical, it's operational. Drawing on work inside regulated, security-conscious enterprises, I'll walk through what changes when an AI system moves from pilot to production: who owns it, what culture does to even a well-written review process, and why governance, done right, is what lets organizations move faster rather than slower. The talk closes with a practical, four-move playbook that any team, not just ones with a dedicated AI platform, can start using immediately.
Bio: Mariano Mattei builds AI agent systems that hold up under audit. He has spent more than thirty years in software engineering and security, and the last several building agentic platforms for clients where a wrong answer has consequences: regulated manufacturers, professional services firms, and organizations running AI against records they have to defend. His career began with twenty-eight years at IBM, as a Senior Software Engineer and then as manager of the company's global SWAT incident response teams, doing root cause analysis on mission-critical systems worldwide. He moved into security leadership as Director of Governance, Risk and Compliance at Layer 8 Security, then as Vice President of Cybersecurity and AI Solutions at Azzur Group, where he built AI-driven platforms inside regulated environments. He founded Mattei Systems to do that work directly: designing, building, and assessing agent systems with verification, audit trail, and human approval built in from the start rather than added after an incident.

Vicki Boykis: Small but Mighty: ONNX + ONNXRT for Model Serving Abstract: Today's machine learning inference landscape discourse is largely made of up of conversations around generative models and inference engines like vLLM and SGLang. But the production model landscape is diverse and 100B+ parameter generative inference patterns (autoregressive decoding and the KV-cache lifecycle) don't account for all use-cases: Some are smaller transformer architectures, some are encoder-only transformers for embedding, ranking, and classification. This talk will cover how ML inference works generally and narrow in on ONNX and ONNX Runtime in production. Bio: I'm currently a founding machine learning engineer working on recsys/personalization/information retrieval. Links: vickiboykis.com/ @vboykis/ @vickiboykis.com on bsky

❗Please note, by RSVPing you agree to comply to our code of conduct - https://github.com/DataPhilly/projects/blob/main/COC.md