​An event dedicated to exploring all things AI Engineering!

​Event partners: WorkOS, You.com & Veris

​​Agenda

​​​​​​🕒 5:30 pm – 5:35 pm
Welcome and Opening Remarks

​🎤 Marijan Cipcic, Principal Events Manager at Daytona

​​​​​​​​🕒 5:35 pm – 5:50 pm
Talk "Teaching a Weak Code Model to Improve with Reinforcement Learning"​

​🎤 Daniel (Thi) Graviet, Machine Learning Engineer at Daytona

​​​​​​​​Outline:

​What happens when you give a small, weak code model a way to learn from its mistakes?

​In this live demo, we'll use Group Relative Policy Optimization (GRPO) to fine-tune a code model through trial and error. The model will generate solutions to coding tasks, run them in isolated Daytona sandboxes, and receive rewards based on whether its code passes the tests.

​We'll walk through the full loop—from generating completions and grading them in parallel, to feeding the rewards back into training and watching the model improve over just a few steps. Along the way, we'll explore why reinforcement learning for code is challenging, how sandboxed execution makes automated feedback possible, and what it takes to make these systems reliable enough for a live stage demo.

​You'll leave with a practical look at:

​- How GRPO trains models without traditional labeled answers
- Using Daytona sandboxes as safe, parallel code graders
- Designing simple, explainable rewards for coding tasks
- Observing a model's performance improve in real time

​​🕒 5:50 pm – 6:05 pm
Talk "Build a Team of AI Specialists - from Slack"

​🎤 Ryan Cooke, Engineer Manager at WorkOS

​​​​​​​​Outline:

​AI tools give everyone the same general-purpose assistant. Atlas lets everyone at WorkOS create specialists and recurring workflows specific to their tasks at hand, directly from Slack.

​In this live demo, Ryan will show real tasks across connected systems, easily turn it into scheduled automation, and create a named AI teammate with its own instructions, memory and restricted tool access.

​This turns successful conversations into durable agent operations, workflows that keep running and specialists anyone can call by name.

​​🕒 6:05 pm – 6:20 pm
Talk "Self-Improving Agents from Behavioral Exhaust"

​🎤 Edward Irby, Staff Software Engineer at You.com

​​​​​​​​Outline:

​An agent that improves itself by reading its own traces. Every event selected, every thread suspended, every deadlock encountered becomes training data for its own behavior — not through weight updates, but through structured iteration across two mutable surfaces: behaviors and skill instructions.

In a Daytona sandbox space, the agent authors candidate behaviors and skills, and a deterministic gate scores each variant from the trace alone — pass, fail. The winners are promoted into a space as plugins; the rest are discarded. The kernel itself never rewrites in place — it grows by composition.

Behavioral programming makes this native: behaviors compose as independent scenarios that the arbiter re-evaluates fresh each step, so adding or removing one never corrupts the rest. Every candidate is verified for deadlocks before it runs — self-modification can't break confluence.

The result is a neuro-symbolic harness: neural generation proposes, symbolic verification disposes — and the exhaust is the teacher.

​​🕒 6:20 pm – 6:35 pm
Talk "Give Your Agent a World, Not a Mock"

​🎤 Ritiz Tambi, Senior AI Engineer at Veris AI

​​​​​​​​Outline:

​Sandboxes hand an agent a computer in seconds. Coding agents are fast, and a fleet turns out a week of changes in an afternoon. Verification is the constraint now. An agent's code calls Stripe, or Postgres, or the CRM, and nothing answers. So it writes a mock, asserts against its own mock, and hands the PR to a human to check. The more services the code touches, the more of it nobody has actually run. Mocks can't close that gap because they hold no state, staging can't because a fleet can't share one, and a vendor's test mode won't fail when you need it to. So we built the other half. Not mocks, a world. Every agent gets its own. A full stack of everything it depends on, where the clock, the failures, and the credentials behave like the real ones. It seeds the history it needs, runs the flow, breaks it on purpose, checks what actually happened. Then it promotes that world so the rest of the fleet starts from something closer to production, catches the bugs that only show up there, and ships PRs that hold.

​​🕒 6:35 pm – 6:45 pm
Talk "How Sandboxes Power a Software Factory"

​🎤 Daria Shifrina, Member of Technical Staff at Obvious

​​​​​​​​Outline:

​You've gotten AI to write code for you. Where does all that code run? At Obvious, thousands of isolated sandboxes turn our product from a fast pair programmer into a software factory. I'll cover how that architecture looks, why we put an OS provider layer underneath it, and where this goes: a sandbox marketplace for your software factory.

​​🕒 6:45 pm – 6:55 pm
Talk "Stop talking to your AI Agent! Teach it instead."

​🎤 Marina Trajkovska, GTM Engineer at Vellum

​​​​​​​​Outline:

​The way we interact with AI agents is broken. You type instructions, they follow them, you repeat yourself. Companion mode changes that. You talk, you type, you demonstrate on your machine while the agent watches your screen. It learns the skill from what you do, not what you say.

​​​​​​​​​​​​​​​​🕒 6:55 pm - 8:30 pm

​​​Networking

​With pizzas and beverages

​About event

​This is dynamic gathering for AI enthusiasts, innovators, and professionals to collaborate, share ideas, and explore the latest advancements in artificial intelligence. Whether you're building AI products, researching cutting-edge algorithms, or simply passionate about the field, join us to connect, learn, and drive the future of AI forward.