One line disappeared from a system prompt during an ordinary week of edits. The tests stayed green. The JSON stayed valid. And the classifier quietly began routing almost every request to the wrong specialist — a production failure nobody caused by touching the code.
This evening is about the failures that look correct: behavioural drift in LLM systems. Output that a human reader would sign off on, that passes every schema check, and that safe fallbacks quietly paper over — so a mis-routed request can go unnoticed for days. We'll look at why the usual answers (add validation, write more evals) don't catch it, and what testing, observability, and plain engineering discipline actually look like when your software is non-deterministic.
Then we open it up: a panel on who owns reliability when the code didn't change — system prompts as code vs. config vs. content, safe fallbacks vs. loud failures, and whether an LLM feature can ever earn the same trust as deterministic code.
Developers and architects running LLM features in production, or about to — this one's for you.
- Schedule
- 18:00 — Doors, drinks & networking
- 18:30 — Welcome & housekeeping
- 18:35 — Talk "Behavioral drift: the silent bugs in LLM workflows" by Andrey Minogin + Q&A
Andrey Minogin is a software architect in Vienna. He has spent seven years building a JVM and PostgreSQL product, most recently LLM workflows in production, and also does fixed-scope performance audits for Java, Kotlin and Postgres systems. He writes at minogin.com.
• 19:15 — Short break
• 19:25 — Panel discussion
\- Martin Galusinski, SQUER
\- Andrey Minogin, software architect, Vienna
\- TBD
• 20:05 — Open networking
Details
• 📅 Tuesday, 6 October 2026
• 🕕 18:00
• 🗣️ Language: English
• 🎓 Level: Intermediate — developers and architects