Guardrailing Vibe Coding | Stockholm MLOps Event #32 | May 5 2026

Insights from the Stockholm MLOps community

At Stockholm MLOps #32, the conversation moved beyond prompts, demos, and AI-assisted code generation.

The focus shifted toward something much more operational:

How do you maintain control over probabilistic systems inside real organizations?

Across talks, demos, and the hands-on hackathon, one signal became increasingly clear:

The hard part is no longer generating code.

The hard part is:

  • defining intent

  • structuring business logic

  • validating outputs

  • enforcing boundaries

  • and operating AI systems safely at runtime

The event repeatedly challenged the idea that “vibe coding” alone is enough for enterprise AI.

Instead, the strongest discussions focused on:

  • specifications

  • runtime governance

  • critique loops

  • structured context

  • human validation

  • and operational guardrails

Summary — Key Insights from the Event

  • Guardrails are moving from policy documents into runtime systems

  • Specifications are increasingly replacing prompts as the source of truth

  • Trust comes from system design, not model intelligence

  • Human validation is becoming a core operational step

  • AI-assisted development is shifting work upstream toward alignment and structure

  • Enterprise AI depends on company knowledge, not just model capability

  • Critique loops and evals are becoming part of production architecture

Key Insights from the Event

Guardrails Are Becoming Runtime Infrastructure

One of the clearest signals from the event was that governance can no longer live only in documentation or compliance frameworks.

Eyo Eyoma from 8wave.ai argued that:

“Governance cannot just be a framework. It has to move into runtime.”

The discussion repeatedly focused on:

  • rule adherence checks

  • evals during execution

  • approval steps

  • runtime monitoring

  • and auditability

The implication was clear:

As AI systems become more autonomous, governance increasingly becomes part of operational infrastructure.

Vibe Coding Breaks When Business Rules Are Hidden

Several talks exposed the limitations of prompt-first development.

Karin Nissfolk emphasized:

“AI cannot guess your company’s rules.”

The strongest operational signal was that enterprise correctness is contextual.

A model may generate plausible output while still violating:

  • internal policy

  • business definitions

  • compliance requirements

  • or operational logic

The event repeatedly showed that:
working code does not automatically mean correct business behavior.

Specifications Are Becoming The Source Of Truth

Qlerify demonstrated a workflow where prompts were treated only as the starting point.

Nikolaus Varzakakous explained:

“Instead of vibe coding with a prompt, you produce a JSON spec with thousands of lines.”

The workflow focused on:

  • domain models

  • business behavior

  • workflows

  • structured context

  • test cases

  • and bidirectional specifications connected through MCP-compatible tooling

The strongest signal was that specifications increasingly matter more than prompts.

Trust Comes From System Design, Not Model Quality

Karin Nissfolk pushed one of the strongest operational themes of the evening:

“This is not a model problem. It’s a business problem.”

The event repeatedly challenged the idea that better models automatically create trustworthy systems.

Instead, trust was framed as a property emerging from:

  • structure

  • validation

  • runtime controls

  • business alignment

  • and human understanding

Or as Karin described it:

“Trust is understanding the output, knowing it is correct in your business context, and that it solves a business problem.”

Human Validation Is Becoming The Deployment Gate

A recurring pattern throughout the event was that code generation itself is increasingly becoming the easy part.

Nikolaus Varzakakous stated:

“The main thing is validating it with humans in the loop.”

The practical bottleneck increasingly shifts toward:

  • validating intent

  • aligning workflows

  • reviewing outputs

  • and ensuring systems reflect operational reality

The event strongly suggested that enterprise AI is becoming less about coding speed and more about organizational alignment.

Critique Loops Are Becoming Part Of Architecture

Another strong signal was the growing importance of critique loops and iterative evaluation.

Eyo Eyoma explained:

“The first output is not good enough.”

And:

“You want the LLM to critique itself, or use a different model to critique the first output.”

The event framed critique not as optional quality improvement, but as part of system reliability itself.

This included:

  • multi-step evaluation

  • cross-model critique

  • rule adherence checks

  • runtime validation

  • and structured feedback loops

Company Knowledge Is The Real AI Dependency

Several discussions rejected the idea that enterprise AI success comes primarily from model capability.

Karin Nissfolk summarized this clearly:

“AI does not replace company knowledge. It depends on it.”

The strongest signal was that organizations succeed when:

  • business rules are explicit

  • workflows are modeled

  • definitions are structured

  • and operational logic is documented

The limitation is often not the model.

It is the organization’s ability to structure its own knowledge.

The Organization Is Becoming The Bottleneck

One of the most repeated themes of the evening was that technical capability is moving faster than organizational change.

Nikolaus Varzakakous stated:

“Individual developer productivity is going crazy, but nothing is happening on the company level.”

And:

“The slowest moving part is the organization.”

The event repeatedly highlighted the widening gap between:

  • what AI systems can technically generate

  • and what organizations can validate, govern, and operationalize

Operational Signals Emerging From Meetup #32

The event reflected several larger shifts happening across AI-assisted development:

  • prompts are becoming specifications

  • coding is becoming orchestration

  • runtime governance is becoming operational infrastructure

  • trust is becoming a system property

  • AI workflows increasingly depend on critique loops and evals

  • enterprise AI requires deterministic boundaries around probabilistic systems

  • human validation is becoming part of deployment architecture

The output of AI-assisted development is increasingly not just code.

It is:

  • specifications

  • workflows

  • tests

  • traceability

  • runtime monitoring

  • and operational control systems

What This Event Signals

Meetup #32 strongly suggested that AI-assisted software development is moving out of the “prompt and pray” phase.

The next phase is operational.

The developer role increasingly shifts toward:

  • structuring intent

  • defining constraints

  • validating workflows

  • orchestrating tools

  • maintaining traceability

  • and designing reliable systems around probabilistic models

The core signal from the event:

Vibe coding is powerful, but without specifications, guardrails, and human validation, it does not become enterprise software.

Speakers & Companies Featured

  • Nikolaus Varzakakous — Qlerify

  • Staffan Palopää — Qlerify

  • Eyo Eyoma — 8wave.ai

  • Karin Nissfolk — AI Risk & Security / Data & Analytics

Topics covered:

  • vibe coding

  • AI governance

  • runtime guardrails

  • specification-driven development

  • agentic workflows

  • AI-assisted software development

  • trustworthy AI systems

  • critique loops

  • operational AI control

Related Insights

Upcoming Events

👉 AI Agents and the Future of Health — Stockholm MLOps #33

Event Details

📍 AI Sweden, Stockholm
📅 May 5, 2026
👥 Stockholm MLOps Community

👉 Meetup Event:
Guardrailing Vibe Coding | Stockholm MLOps Event #32

👉 Our Meetup Community:

Previous
Previous

AI Agents and the Future of Health | Stockholm MLOps Event #33 | May 18 2026

Next
Next

Sovereign AI in Production | Stockholm MLOps Event #31 | April 23 2026