Local-first agents
AI agents that run on hardware people already own. Private by architecture, not by promise. It's the foundation OOMU is built on.
Research
Applied R&D. Emphasis on applied.
We don't publish papers and wait. We build tools for our own work, use them daily, and turn the ones that prove themselves into products. OOMU is the first to make that trip.
Our method
Everything we ship, we used first.
Every Eldris experiment starts as internal tooling with a real job to do. If it makes our own work faster, safer, or cheaper for months on end, it earns the investment to become something others can use. If it doesn't, we learn and move on.
Research areas
Where the lab spends its time.
Hybrid model economics
Routing work between local and frontier models so cost tracks the difficulty of the task, not the number of keystrokes. Paying for the cloud only when the cloud earns it.
AI security & safety
Where Eldris began. Agent containment, permissioning, human-approval design, and the audit trails that make autonomous systems accountable.
Human-legible workflows
Making agent behavior visible, editable, and reviewable by the people it works for. Trust starts with understanding.
Case in point
OOMU: from lab bench to launch.
OOMU began as the lab's own workstation: a way to run agent workflows on our own machines without surrendering privacy or paying cloud rates for routine work. It combined three research threads at once: local-first agents, hybrid model routing, and human-review design.
Months of daily internal use shaped it into something worth sharing. Today it's our first mass-market product, and the clearest expression of how the lab works.
Collaborate
Have a hard, practical problem?
We partner with organizations, universities, and research groups on applied AI problems where the answer has to work outside the lab. If you're wrestling with agent deployment, privacy-preserving AI, or AI security, we'd like to hear about it.
Research that has to work on Monday.
Eldris
Applied AI, built on trust.