DSL-based Safety Guardrails for LLM Applications

June 2026

LLM-based applications need safety checks, for example on prompts, generated outputs, or tool calls. Today, these checks - if done so - are implemented in ad hoc if a problem is detected. This makes them hard to reuse, maintain, and reason about.

Tasks:

Requirements Analysis: Analyze requirements for runtime safety guardrails in LLM applications.
DSL Design: Design a domain-specific language (DSL) for specifying guardrails.
Implementation: Implement a Python library for request interception and guardrail enforcement.
Enforcement Actions: Define and support a set of enforcement actions such as block, rewrite, retry, and log.
Evaluation: Evaluate the approach using representative example use cases.

Thesis Goal The goal of this thesis is to design and implement a Python library, inspired by LMQL, that provides a small DSL for specifying safety guardrails. LLM requests should be routed through the library, which evaluates and enforces the guardrails at runtime.

Supvervision

This thesis will be co-supervised by Philipp Zech and Michael Vierhauser.