[talks] Pydata Lausanne Sept 2026 @ Visium

The event will be kindly hosted by Visium.
Please register using the form :
https://forms.gle/XKTFDoUdnUqAoxNQ7

👇 👇 Line-up 👇👇

AgentOps — The Next Layer After MLOps: Managing, Deploying and Monitoring AI Systems
by Andy Bermond (VISIUM)

LangChain's 2026 State of Agent Engineering survey put 57% of respondents with agents in production, 89% with observability in place — and only 52% running evals. Quality was the top barrier, named by nearly a third. Tracing has become table stakes; knowing whether the thing is any good has not. That gap is the interesting part.

The term AgentOps was coined in a 2024 paper out of CSIRO's Data61. Its argument: existing tooling mostly tracks LLM-specific metrics and prompt management, with little support for the artifacts that actually define an agent — its goals, its plans, its tools.
MLOps grew out of a similar observation: the model was never the hard part, the system around it was. Agents make the point again, louder. A model is a function — you version it, test it, roll it back. An agent is a process. It holds state, calls tools, spends money, retries, and occasionally decides something nobody specified.
So what is the deployable unit now? What does a release look like when behaviour changes without the artifact changing? What do you monitor when there is no accuracy number, and what does a rollback mean when the failure is a loop of tool calls rather than a bad prediction?

Eval-Driven Development for LLM Systems
by Lucas Eckes (VISIUM)

Testing traditional software and ML systems is reliable work. We assert: same input, same output, a hand-written expected value on the right-hand side, and a suite that still passes in five years.

So what happens when the feature under test is a language model? The same request returns a different string every run, all of them potentially correct. Ground truth rarely arrives because a user who gets a bad answer closes the tab rather than telling you. And the artifact that changed is a prompt, with no hash, no version bump, and no retraining run.

This is a talk about what replaces the suite, at the level of a single application. Where your test cases come from before you have production traffic, and why the ones you write on day three stop describing the system by month six. Which checks need a model to evaluate and which collapse back into plain Python. How to calibrate an LLM judge before you trust the numbers it produces.

Demonstrated end-to-end on a small application with Arize Phoenix, open source and local

Give your agent a local sandbox, not just tools
by Thibault Viglino (VISIUM)

LLMs are often faster, cheaper and more reliable when they write a short program that calls your tools instead of making a sequence of individual tool calls.

Pydantic Monty runs isolated Python with no container, VM or sandboxing service in the loop. Creating a sandbox and running ten commands in it takes 5 ms (Docker: 900 ms, a sandboxing service: 1900 ms).

In this talk we will motivate why a local sandbox is useful and showcase it with Monty. We will see how it interacts with the existing tooling and MCP ecosystems.

TL;DR: Nobody Reads Your Agent's Output
by Christian Abbet (VISIUM)

We hand more and more of our coding work to agents, and we consume what they produce everywhere: code review, documentation, and issue reports. Every run returns a wall of text that no human wants to read (and that machines don't need either) as most of it is often repetition or nitpicks.

This talk asks whether tooling can fix that: output that is short, sharp, and actually worth reading. Less text generated now means less text ingested by every downstream pipeline.

Or maybe one output was never enough. A second option is to have the agent emit different messages: adding context the agent needs and the human would find tedious, or dropping framing the human wants and the agent pays tokens for. Each reader gets what it needs, and neither pays for the other's format.

--- --- --- --- --- --- --- --- --- ---

NOTES

• PLEAS RSVP: If your status changes, change your RSVP accordingly! This enables us to manage reservations.
• If you are on the waiting list: we will do our best to increase the number of seats, so stay tuned