About
What is Relaxed Constraints?
Relaxed Constraints is an autonomous research lab where humans and AI agents collaborate to build self-improving systems. We design our processes, workflows, and culture around the capabilities of AI agents rather than retrofitting automation onto human processes. The lab itself is the experiment: a self-improving digital organism that gets better at what it does with every iteration.
The name Relaxed Constraints reflects how we operate. We relax the constraints of traditional organisations — rigid hierarchies, manual approvals, human-speed decision making — and replace them with agent-driven processes under human oversight. We also relax the constraints of fixed schemas in our ML research, working with data that is messy, nested, and inter-connected.
What We Do
Autonomous agents. We build AI agent systems that plan multi-step workflows, evaluate their own outputs, and operate with minimal human intervention. Every process that can be delegated to an agent is. Humans drive prioritisation, design decisions, and interpretation — the mechanical work is handled by agents.
Self-improving systems. Every agent action, experiment run, and routing decision produces a trace. These traces are themselves data: analysed by other agents, used to improve processes, and fed back into the system. The lab learns from its own behaviour and gets better over time.
ML & data systems. We develop neural network architectures for semi-structured data — density estimation, representation learning, and generative models that work natively with hierarchical JSON-like structures. We apply these to practical database problems: cardinality estimation, query optimisation, learned indexes, and synthetic data generation for document databases.
Background
Relaxed Constraints was founded by Thomas Rückstieß, who holds a PhD in Machine Learning from the Technical University of Munich and spent 13 years at MongoDB, most recently as Head of Machine Learning Research.
We have an ongoing collaboration with the University of Sydney, where Thomas works with the databases group and supervises PhD and Honours students working on learned query optimization, index recommendation, and AI-assisted experimentation.
Principles
Agents are first-class citizens. We use AI agents unapologetically — for content creation, code reviews, research, and decision making. They’re not assistants waiting for instructions; they’re active participants in how the lab operates. If an agent can do it well, we let it.
No task is human-only. Every task in the lab can be done by a human, an agent, or both working together. Every resource (code, data, documentation, website, infrastructure) is accessible to AI. This isn’t about replacing humans; it’s about removing the artificial boundary between what humans can touch and what agents can touch, so work flows to whoever or whatever is best suited for it.
Everything is observable. Every agent action, every experiment run, every routing decision produces a trace. These traces are measured, analysed, and fed back into the system — so the lab learns from its own behaviour and improves over time. Nothing is a black box.
Get Involved
While we’re not currently hiring, we’re always looking to connect with researchers, students, and engineers who share these interests. We value curious and self-motivated people who want to exchange ideas, challenge assumptions, and learn from each other. If that sounds like you, get in touch.