The open-source MLOps platform
Where engineers and agents work together.
You set the objective. Agents run the experiments. Every run is tracked and reproducible, one click from production, and monitored once it's there.
The Problem:
Everyone wants agents doing ML work. Nobody has a clear structure.
The capability isn't the gap. Agents already write the training code, run the experiments, and can carry a model the whole way to production.
What's missing is the structure around that — where runs get tracked, how they get compared, explained, reproduced, and productionized. So every team improvises: different ways of working, different tools, wired to agents differently each time.
It holds up until you have to explain a result, rerun it, or hand it to someone else.
The Solution:
We have a structure. It is called LUML.
Five steps, from your first instruction to a model running in production.
The Solution:
We have a structure. It is called LUML.
Five steps, from your first instruction to a model running in production.
STEP 01 / 05
Work with agents, side by side
You drive a coding agent from the terminal you already use. What it produces appears beside you as work you can look at: experiments, metrics, models. Whether you trained it or the agent did, it is captured the same way.
STEP 02 / 05
The agents get better at your domain
None of that session is thrown away. What worked, what you overruled, and the details specific to your domain get consolidated into your team's context — so the next session starts from the knowledge you already provided, instead of from scratch.
STEP 03 / 05
Hand over the objective
Once the context holds enough, you can stop supervising each run and state a goal instead. Prisma forks it into parallel branches, runs each in its own git worktree and proposes the next direction. Weak branches are pruned, a fixed metric defines the winner, and the winning model comes back to you for review.
STEP 04 / 05
The winning model folds into one artifact
Weights, metadata, dependencies and preprocessing fold into a single self-contained .luml artifact. The registry versions it and links it back to the exact experiment that produced it. Nothing about how it was made is left behind on a laptop.
STEP 05 / 05
Deployed and monitored
A service you host inside your own network reaches out, pulls the artifact and serves it, so nothing has to reach in. Monitoring sits on the same machine, watching drift, data quality and health where the model actually runs.
Meet the Team
We built the tool we always needed
We're data scientists and engineers who've built AI systems end-to-end at L'Oréal · Bitpanda · SCCH — across startups and large-scale environments.
FAQ
Frequently
asked
questions
The path from first experiment to production, for humans and agents on one platform: experiment tracking with traces and evals, autonomous research loops, and a registry with deployments and monitoring on your own infrastructure.