Vstorm at Py AI: agentic AI in production with OpenAI, SurrealDB and Theory Ventures

On March 10, 2026, Vstorm CEO Antoni Kozelski joins Jason Liu of OpenAI, Tobie Morgan Hitchcock of SurrealDB, and host Bryan Bischof of Theory Ventures on a panel at the Py AI Conference in San Francisco. The panel focuses on what it takes to run agentic AI in production and the reliability, observability, and architecture decisions that determine whether a system survives first contact with real users. This article will introduce the event, the panellists, and what Vstorm brings to the table.

What Py AI is and why it matters

Py AI is a community built by the teams behind Pydantic and Prefect for one specific audience: Python developers who are shipping AI to production, not just preparing demos.

The conference originated from a straightforward observation. Most AI events are built around what AI can theoretically do. Py AI was built around what practitioners actually face once a system is running, tackling subjects like reliability at scale, iteration without regression, and the operational complexity that appears the moment real users arrive.

“The hard part was never getting a demo to run. It was everything after: reliability at scale, iteration without regression, the thousand small fires that ignite the moment real users show up.”
– Py AI founders, Samuel Colvin, CEO of Pydantic, and Adam Azzam, VP Product of Prefect

This aligns well with what we at Vstorm have observed across 30+ agentic AI implementations, that the gap between a working prototype and a system that provides real results in production is where most projects fall short.

The panel: who is in the room

The March 10 panel brings together practitioners who will each present a distinct piece of the challenge around the successful implementation of production-grade agentic AI solutions.

Antoni Kozelski: CEO & Co-founder of Vstorm

Antoni leads Vstorm’s mission to apply agentic AI to mission-critical business processes, providing his knowledge and insights gained from practical applications of successful AI projects in different sectors. Additionally, he will be providing a lighting talk on the Logfire AI Assistant later that day, a practical implementation built on the Pydantic stack that Vstorm contributed to, demonstrating how observability tooling can be embedded directly into agentic workflows.

Jason Liu: Consultant and Instructor at OpenAI

Jason is the creator of the Instructor library, the widely adopted tool for structured outputs from language models. His work directly addresses one of the most common causes of agent failure in production: unpredictable, unvalidated model outputs.

Tobie Morgan Hitchcock: CEO & Co-Founder, SurrealDB

Tobie founded SurrealDB to solve a problem that sits at the foundation of every agentic system: the data layer. SurrealDB unifies graph, document, vector, time-series, and real-time data in a single engine. With 17+ years in distributed systems, Tobie’s focus is on durable agent memory, context precision, and architectures that scale beyond single-writer databases.

With Bryan Bischof as host: Head of AI, Theory Ventures

Bryan brings a practitioner’s lens to the panel, with a PhD in mathematics, co-author of an O’Reilly book on production recommendation systems, former Head of AI at Hex where he led the team building their analytics copilot, and currently teaching data science at Rutgers graduate school.

The challenge they will address: what production-grade AI actually means

Most teams approaching agentic AI attempt to manage production problems with fragmented tooling, such as separate observability stacks or manual debugging of agent failures, often with no systematic method for tracing decisions through multi-step pipelines. With issues being patched reactively once they surface in live systems. The panel will discuss what separates the pilot stage from a production system that operation teams can count on.

The panel will examine the building blocks of reliable Agentic AI systems:

• Structured outputs: preventing unpredictable model behaviour at the point of output generation
• Data architecture: giving agents durable memory and precise context retrieval at scale
• Observability: tracing agent decisions in production so teams can debug, audit, and improve
• Investment and system-level thinking: evaluating which architectural decisions deliver lasting value

Together, these topics form a practical checklist for any team trying to move an agentic system from concept to production.

What Vstorm brings to the table: 30+ practical implementations

Vstorm brings concrete application experience to this conversation. We do not study agentic AI from a distance, we build business critical production systems for mid-market companies and have done so across healthcare, e-commerce, engineering software, and investigative journalism, among others.

The following two case studies illustrate the operational reality the panel will discuss:

AI multi-agent for order completion for PoD leader, Mixam

Mixam operates a complex print-on-demand platform where customers navigate hundreds of product configurations. Previously, their support team handled this manually, with significant friction and inconsistency. Vstorm designed and deployed a multi-agent system that guides customers through the full order process. The result: an 11.76% increase in orders and a 95.4% success rate in workflow completion. You can read more about Mixam’s agentic system here.

Multi-channel pre-appointment agent for healthcare

A U.S. healthcare provider serving 100,000+ members was managing patient scheduling and pre-appointment communication entirely through manual nurse and administrative workflows. After deploying a multi-channel agentic system, each doctor now saves more than five hours per week and patient engagement has risen over 20%. More can be found in this complete case study.

These are not pilot outcomes. They are production systems running on live infrastructure.

Vstorm is also proud to be the first tech consulting company accepted as a member of the Agentic AI Foundation and a contributing partner of Pydantic AI, making this panel a natural extension of our consultation and engineering work.

Why this panel is relevant to mid-market operators

If you are a COO, Head of Operations, or CTO at a mid-market company evaluating agentic AI, the Py AI panel addresses the questions that matter most once an initial pilot shows promise: What does it cost to maintain this in production? How do we trace failures? What data architecture do we need to give agents reliable memory?

The conversation beginning at 1:50pm on March 10 will offer concrete answers from engineers who have built and deployed real working agentic systems.

The Py AI Conference takes place March 10, 2026 in San Francisco. You can register at pyai.events.